Models / Changes
The model at that id is not the model you tested.
238 of 409 tracked open-weight models changed a tracked key after their first upload; 124 changed a structural one, 69 of them more than a week after release. Each row is one commit that changed a model's config, tokenizer, chat template, generation defaults or weights, with what that change does.
Models with history
409
Changes
586
Structural
188
Tokenizer / template
345
By most consequential class: 188 structural, 185 tokenizer, 160 template, 29 generation-defaults, 0 weights-only, 24 metadata. 313 of the 586 landed more than a week after the model's first commit, so they are not release-day fixes. History reaches back to 2019-02-18; 0 of the changes were caught by a daily snapshot rather than recovered from git. Last read 2026-10-18.
| Date | Model | Class | What changed, and what it does |
|---|---|---|---|
| 2026-10-14 | zai-org/GLM-5.3 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-10-14). |
| 2026-10-14 | baidu/Unlimited-OCR | structural template | model_type "unlimited-ocr" to "unlimited_ocr"; a different class is instantiated from the same id; a chat template was added; apply_chat_template now formats prompts that it previously refused or formatted by default (2026-10-14). |
| 2026-10-05 | unsloth/gpt-oss-20b-BF16 | tokenizer template | EOS changed from [200002,199999] to [200002,199999,200012]; anything fine-tuned or served against the old value stops at a different token; the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-10-05). |
| 2026-09-29 | openbmb/MiniCPM5-2B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-09-29). |
| 2026-09-14 | biohub/ESMC-6B | structural tokenizer | EOS declared as 2 (was unset); generation that relied on the library default now stops on this token; BOS declared as 0 (was unset); head dim unset to 64, MLP width unset to 6912, attention heads unset to 40, layer count unset to 80, KV heads unset to 40 (2026-09-14). |
| 2026-09-05 | CMSManhattan/JiRackUltra_1b | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-09-05). |
| 2026-09-04 | zai-org/GLM-5.3 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-09-04). |
| 2026-09-01 | farbodtavakkoli/OTel-2.0-LLM-31B-IT | structural tokenizer template | EOS changed from 106 to [1,106]; anything fine-tuned or served against the old value stops at a different token; derived parameters 30.15B to 30.56B (+418.7M, +0.84 GB of bf16 weights); architectures declared as ["Gemma4ForConditionalGeneration"] (was unset) (2026-09-01). |
| 2026-08-29 | ornith-ai/Ornith-1.5-397B-NVFP4 | structural | quantization_config changed; the stored format of the weights is different (2026-08-29). |
| 2026-08-28 | RadixArk/Qwen3.8-27B-DSpark | structural tokenizer | EOS changed from 248046 to 248044; anything fine-tuned or served against the old value stops at a different token; BOS 248044 removed; MLP width 10240 to 17408, attention heads 40 to 32 (2026-08-28). |
| 2026-08-24 | ai-sage/Giga-Embeddings-instruct-480M-0826 | structural | context window 32768 to 8192 tokens (2026-08-24). |
| 2026-08-23 | OBLITERATUS/Qwen3.8-27B-OBLITERATED | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-08-23). |
| 2026-08-23 | OBLITERATUS/Qwen3.8-27B-OBLITERATED | tokenizer | EOS changed from 248044 to 248046; anything fine-tuned or served against the old value stops at a different token; pad token unset to 248044; batched training masks a different id (2026-08-23). |
| 2026-08-23 | OBLITERATUS/Qwen3.8-27B-OBLITERATED | tokenizer | EOS changed from [248046,248044] to [248046]; anything fine-tuned or served against the old value stops at a different token (2026-08-23). |
| 2026-08-22 | OBLITERATUS/Qwen3.8-27B-OBLITERATED | structural tokenizer template | derived parameters 26.79B to 26.79B (+7.0M, +0.01 GB of bf16 weights); architectures ["Qwen3_5ForCausalLM"] to ["Qwen3_5ForConditionalGeneration"]; a different class is instantiated from the same id; the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-08-22). |
| 2026-08-22 | OBLITERATUS/Qwen3.8-27B-OBLITERATED | structural tokenizer | derived parameters 26.79B to 26.79B (-7.0M, -0.01 GB of bf16 weights); architectures ["Qwen3_5ForConditionalGeneration"] to ["Qwen3_5ForCausalLM"]; a different class is instantiated from the same id; add_bos_token false to unset; encoding the same text yields different ids (2026-08-22). |
| 2026-08-21 | bloomer010/Ling-3.0-tiny-GGUF | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-08-21). |
| 2026-08-20 | OBLITERATUS/Qwen3.8-27B-OBLITERATED | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-08-20). |
| 2026-08-19 | OBLITERATUS/Qwen3.8-27B-OBLITERATED | structural tokenizer template | derived parameters 26.79B to 26.79B (+7.0M, +0.01 GB of bf16 weights); architectures ["Qwen3_5ForCausalLM"] to ["Qwen3_5ForConditionalGeneration"]; a different class is instantiated from the same id; a chat template was added; apply_chat_template now formats prompts that it previously refused or formatted by default (2026-08-19). |
| 2026-08-19 | ornith-ai/Ornith-1.5-397B-FP8 | structural | quantization_config changed; the stored format of the weights is different (2026-08-19). |
| 2026-08-19 | ornith-ai/Ornith-1.5-35B-A3B-FP8 | structural | quantization_config changed; the stored format of the weights is different (2026-08-19). |
| 2026-08-19 | ornith-ai/Ornith-1.5-35B-A3B | tokenizer | add_bos_token unset to false; encoding the same text yields different ids (2026-08-19). |
| 2026-08-13 | webAI-Official/TwIL-LM3 | tokenizer | tokenizer model_max_length 131072 to 65536; truncation defaults change; tokenizer_class unset to "PreTrainedTokenizerFast" (2026-08-13). |
| 2026-08-10 | nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 | structural | context window 262144 to 1048576 tokens (2026-08-10). |
| 2026-08-05 | nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 | template | derived parameters 32.91B to 105.10B (+72.19B, +144.37 GB of bf16 weights); the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-08-05). |
| 2026-08-05 | nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-08-05). |
| 2026-08-01 | poolside/Laguna-S-2.1-NVFP4 | structural | context window 262144 to 1048576 tokens; quantization_config changed; the stored format of the weights is different (2026-08-01). |
| 2026-07-31 | RadixArk/Kimi-K3-DSpark | structural | rope_scaling added ({"factor":16,"original_max_position_embeddings":65536,"ro...); positions are now scaled (2026-07-31). |
| 2026-07-29 | poolside/Laguna-XS-2.1 | generation-defaults | generation defaults changed (top_k unset to 20); a call that does not pass these explicitly now generates differently (2026-07-29). |
| 2026-07-24 | k-chirkunov/gemma4-e4b-claims-comparison | tokenizer template | pad token "<eos>" to "<pad>"; batched training masks a different id; the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-07-24). |
| 2026-07-23 | ibm-research/MoLFormer-XL-both-10pct | tokenizer | EOS declared as "<eos>" (was unset); generation that relied on the library default now stops on this token; BOS declared as "<bos>" (was unset); tokenizer model_max_length 1e+30 to unset; truncation defaults change (2026-07-23). |
| 2026-07-22 | poolside/Laguna-S-2.1-NVFP4 | structural | quantization_config changed; the stored format of the weights is different (2026-07-22). |
| 2026-07-21 | poolside/Laguna-S-2.1-NVFP4 | template generation-defaults | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering; generation defaults changed (max_new_tokens 32768 to unset); a call that does not pass these explicitly now generates differently (2026-07-21). |
| 2026-07-21 | poolside/Laguna-S-2.1 | template generation-defaults | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering; generation defaults changed (max_new_tokens 32768 to unset); a call that does not pass these explicitly now generates differently (2026-07-21). |
| 2026-07-21 | poolside/Laguna-S-2.1 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-07-21). |
| 2026-07-20 | tencent/Hy3 | declared dtype unset to "bfloat16" (2026-07-20). | |
| 2026-07-14 | microsoft/phi-4 | tokenizer | EOS changed from 100265 to [100257,100265]; anything fine-tuned or served against the old value stops at a different token (2026-07-14). |
| 2026-07-13 | ibm-granite/granite-vision-4.1-4b | structural | vocab_size changed from 100353 to 100352; the embedding and output head change shape, so adapters and merges made against the old vocabulary no longer load cleanly; derived parameters 3.32B to 3.32B (-2.6K, -0.00 GB of bf16 weights) (2026-07-13). |
| 2026-07-12 | ibm-granite/granite-vision-4.1-4b | tokenizer | tokenizer_class "GPT2Tokenizer" to "PreTrainedTokenizerFast" (2026-07-12). |
| 2026-07-08 | OpenMOSS-Team/MOSS-Transcribe-Diarize | tokenizer | add_bos_token false to unset; encoding the same text yields different ids (2026-07-08). |
| 2026-07-07 | poolside/Laguna-XS-2.1 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-07-07). |
| 2026-07-06 | OpenMOSS-Team/MOSS-Transcribe-Diarize | structural | context window 40960 to 131072 tokens; declared dtype unset to "bfloat16" (2026-07-06). |
| 2026-07-05 | tencent/Hy3 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-07-05). |
| 2026-07-03 | Bahushruth/Qwen3.6-35B-A3B-abliterated-v4 | structural | architectures ["Qwen3_5MoeForConditionalGeneration"] to ["Qwen3_5MoeForCausalLM"]; a different class is instantiated from the same id (2026-07-03). |
| 2026-07-03 | Bahushruth/Qwen3.6-35B-A3B-abliterated-v4 | structural | architectures ["Qwen3_5MoeForCausalLM"] to ["Qwen3_5MoeForConditionalGeneration"]; a different class is instantiated from the same id (2026-07-03). |
| 2026-06-30 | nvidia/Qwen3.6-27B-NVFP4 | structural | quantization_config changed; the stored format of the weights is different (2026-06-30). |
| 2026-06-27 | google/tipsv2-so400m14 | structural tokenizer | hidden size unset to 1152, MLP width unset to 4304, attention heads unset to 16, layer count unset to 27; context window unset to 64 tokens; pad token unset to 0; batched training masks a different id (2026-06-27). |
| 2026-06-27 | google/tipsv2-b14 | structural tokenizer | hidden size unset to 768, MLP width unset to 3072, attention heads unset to 12, layer count unset to 12; context window unset to 64 tokens; pad token unset to 0; batched training masks a different id (2026-06-27). |
| 2026-06-25 | ornith-ai/Ornith-1.0-397B-FP8 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-06-25). |
| 2026-06-24 | nvidia/MiniMax-M3-NVFP4 | tokenizer | EOS declared as "[e~[" (was unset); generation that relied on the library default now stops on this token; BOS declared as "]~b]" (was unset); pad token unset to "]!p~["; batched training masks a different id (2026-06-24). |
| 2026-06-19 | z-lab/Qwen3.6-35B-A3B-DFlash | structural tokenizer | EOS changed from 248046 to 248044; anything fine-tuned or served against the old value stops at a different token; layer count 8 to 6, KV heads 4 to 8; derived parameters 1.47B to 1.37B (-100.7M, -0.20 GB of bf16 weights) (2026-06-19). |
| 2026-06-15 | moonshotai/Kimi-K2.7-Code | template generation-defaults | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering; generation defaults changed (temperature unset to 1, top_p unset to 0.95); a call that does not pass these explicitly now generates differently (2026-06-15). |
| 2026-05-27 | poolside/Laguna-XS.2 | generation-defaults | generation defaults changed (max_new_tokens 2048 to 4096, min_p unset to 0, temperature 0.7 to 1, top_p 0.9 to 1); a call that does not pass these explicitly now generates differently (2026-05-27). |
| 2026-05-26 | poolside/Laguna-XS.2 | structural | context window 131072 to 262144 tokens (2026-05-26). |
| 2026-05-12 | moonshotai/Kimi-K2.6 | tokenizer | tokenizer_class "TikTokenTokenizerFast" to "TikTokenTokenizer" (2026-05-12). |
| 2026-05-11 | moonshotai/Kimi-K2.6 | tokenizer | tokenizer_class "TikTokenTokenizer" to "TikTokenTokenizerFast" (2026-05-11). |
| 2026-05-08 | XiaomiMiMo/MiMo-V2.5 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-05-08). |
| 2026-05-07 | XiaomiMiMo/MiMo-V2.5 | tokenizer | EOS changed from [151643,151645,1561672] to [151643,151645,151672]; anything fine-tuned or served against the old value stops at a different token (2026-05-07). |
| 2026-05-07 | nvidia/Gemma-4-26B-A4B-NVFP4 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-05-07). |
| 2026-05-05 | zeroentropy/zerank-2-reranker | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-05-05). |
| 2026-05-01 | nvidia/Gemma-4-31B-IT-NVFP4 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-05-01). |
| 2026-04-30 | moonshotai/Kimi-K2.5 | structural | quantization_config changed; the stored format of the weights is different (2026-04-30). |
| 2026-04-30 | moonshotai/Kimi-K2.6 | structural | quantization_config changed; the stored format of the weights is different (2026-04-30). |
| 2026-04-29 | sakamakismile/Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-04-29). |
| 2026-04-28 | ibm-granite/granite-vision-4.1-4b | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-04-28). |
| 2026-04-28 | XiaomiMiMo/MiMo-V2.5 | structural template | context window 262144 to 1048576 tokens; rope_theta 5000000 to 10000000; every position embedding changes; the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-04-28). |
| 2026-04-28 | nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 | tokenizer | pad token unset to "<|im_end|>"; batched training masks a different id (2026-04-28). |
| 2026-04-27 | XiaomiMiMo/MiMo-V2.5 | tokenizer template generation-defaults | EOS changed from [151643,151645] to [151643,151645,1561672]; anything fine-tuned or served against the old value stops at a different token; the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering; generation defaults changed (temperature unset to 1, top_p unset to 0.95); a call that does not pass these explicitly now generates differently (2026-04-27). |
| 2026-04-21 | nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-04-21). |
| 2026-04-20 | nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 | generation-defaults | generation defaults changed (do_sample unset to true, max_new_tokens unset to 16384, repetition_penalty unset to 1, temperature unset to 0.6); a call that does not pass these explicitly now generates differently (2026-04-20). |
| 2026-04-20 | nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 | structural | architectures ["Nemotron_3_Nano_Omni_30B_A3B_Reasoning"] to ["NemotronH_Nano_Omni_Reasoning_V3"]; a different class is instantiated from the same id (2026-04-20). |
| 2026-04-20 | MiniMaxAI/MiniMax-M2.7 | structural | context window 196608 to 204800 tokens (2026-04-20). |
| 2026-04-18 | nvidia/Gemma-4-31B-IT-NVFP4 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-04-18). |
| 2026-04-17 | z-lab/Qwen3.6-35B-A3B-DFlash | structural | rope_scaling rewritten in a different format ({"beta_fast":32,"beta_slow":1,"factor":64,"original_max_p... to {"beta_fast":32,"beta_slow":1,"factor":64,"original_max_p...); a loader that only reads the old format rejects or ignores the new one (2026-04-17). |
| 2026-04-16 | zai-org/GLM-5.1 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-04-16). |
| 2026-04-10 | nvidia/llama-nemotron-rerank-1b-v2 | structural | rope_scaling rewritten in a different format ({"factor":32,"high_freq_factor":4,"low_freq_factor":1,"or... to {"factor":32,"high_freq_factor":4,"low_freq_factor":1,"or...); a loader that only reads the old format rejects or ignores the new one (2026-04-10). |
| 2026-04-09 | MiniMaxAI/MiniMax-M2.7 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-04-09). |
| 2026-04-09 | MiniMaxAI/MiniMax-M2.7 | declared dtype unset to "bfloat16" (2026-04-09). | |
| 2026-04-09 | jinaai/jina-reranker-m0 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-04-09). |
| 2026-04-07 | nomic-ai/nomic-embed-text-v1.5 | structural tokenizer | head dim unset to 64; derived parameters 108.9M to 110.1M (+1.2M, +0.00 GB of bf16 weights); context window 8192 to 2048 tokens (2026-04-07). |
| 2026-04-07 | nomic-ai/nomic-embed-text-v1 | structural tokenizer | head dim unset to 64; derived parameters 108.9M to 114.8M (+5.9M, +0.01 GB of bf16 weights); rope_scaling added ({"factor":2,"rope_type":"dynamic"}); positions are now scaled (2026-04-07). |
| 2026-04-02 | nvidia/Gemma-4-31B-IT-NVFP4 | tokenizer | pad token "<eos>" to "<pad>"; batched training masks a different id (2026-04-02). |
| 2026-03-30 | nvidia/llama-nemotron-embed-1b-v2 | structural | rope_scaling rewritten in a different format ({"factor":32,"high_freq_factor":4,"low_freq_factor":1,"or... to {"factor":32,"high_freq_factor":4,"low_freq_factor":1,"or...); a loader that only reads the old format rejects or ignores the new one (2026-03-30). |
| 2026-03-16 | nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 | generation-defaults | generation defaults changed (temperature unset to 1, top_p unset to 0.95); a call that does not pass these explicitly now generates differently (2026-03-16). |
| 2026-03-11 | zai-org/GLM-5 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-03-11). |
| 2026-03-09 | stepfun-ai/Step-3.5-Flash | structural | derived parameters 8.28B to 8.81B (+528.1M, +1.06 GB of bf16 weights); tie_word_embeddings unset to false; the output head is no longer shared with the embedding (2026-03-09). |
| 2026-03-06 | sarvamai/sarvam-30b | structural | context window 32768 to 131072 tokens (2026-03-06). |
| 2026-03-04 | nvidia/Kimi-K2.5-NVFP4 | structural | quantization_config changed; the stored format of the weights is different (2026-03-04). |
| 2026-03-04 | sarvamai/sarvam-30b | structural | context window 131072 to 32768 tokens (2026-03-04). |
| 2026-02-27 | amd/DeepSeek-R1-MXFP4 | structural | quantization_config changed; the stored format of the weights is different (2026-02-27). |
| 2026-02-19 | jinaai/jina-embeddings-v5-text-nano-retrieval | tokenizer | tokenizer_class "TokenizersBackend" to "PreTrainedTokenizer" (2026-02-19). |
| 2026-02-16 | jinaai/jina-embeddings-v5-text-nano-retrieval | structural tokenizer | rope_scaling removed (was {"factor":1,"rope_type":"linear"}); positions are no longer scaled; tokenizer_class "PreTrainedTokenizerFast" to "TokenizersBackend" (2026-02-16). |
| 2026-02-16 | jinaai/jina-embeddings-v5-text-nano-retrieval | structural | model_type "jina_embeddings_v5" to "eurobert"; a different class is instantiated from the same id (2026-02-16). |
| 2026-02-12 | MiniMaxAI/MiniMax-M2.5 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-02-12). |
| 2026-02-11 | jinaai/jina-embeddings-v5-text-nano-retrieval | structural | architectures ["EuroBertForMaskedLM"] to ["EuroBertModel"]; a different class is instantiated from the same id (2026-02-11). |
| 2026-02-10 | jinaai/jina-embeddings-v5-text-nano-retrieval | structural | architectures ["EuroBertModel"] to ["EuroBertForMaskedLM"]; a different class is instantiated from the same id (2026-02-10). |
| 2026-02-02 | jinaai/jina-embeddings-v5-text-small | structural | derived parameters 596.0M to 483.0M (-113.0M, -0.23 GB of bf16 weights); context window 40960 to 32768 tokens; model_type "qwen3" to "jina_embeddings_v5"; a different class is instantiated from the same id (2026-02-02). |
| 2026-02-02 | jinaai/jina-embeddings-v5-text-nano | structural | rope_theta 250000 to 1000000; every position embedding changes (2026-02-02). |
| 2026-01-29 | moonshotai/Kimi-K2.5 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2026-01-29). |
| 2026-01-28 | sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 | tokenizer | tokenizer_class unset to "PreTrainedTokenizerFast" (2026-01-28). |
| 2025-12-15 | nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-12-15). |
| 2025-12-15 | nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-12-15). |
| 2025-12-10 | allenai/Olmo-3-7B-Think | generation-defaults | generation defaults changed (do_sample unset to true); a call that does not pass these explicitly now generates differently (2025-12-10). |
| 2025-12-08 | MiniMaxAI/MiniMax-M2 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-12-08). |
| 2025-12-06 | nvidia/Qwen2.5-VL-7B-Instruct-NVFP4 | structural | quantization_config changed; the stored format of the weights is different (2025-12-06). |
| 2025-12-03 | Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8 | structural tokenizer | BOS declared as 151643 (was unset); MLP width 5472 to 6144; quantization_config changed; the stored format of the weights is different (2025-12-03). |
| 2025-12-03 | Qwen/Qwen3-Coder-30B-A3B-Instruct | structural tokenizer | BOS declared as 151643 (was unset); MLP width 5472 to 6144 (2025-12-03). |
| 2025-12-01 | deepseek-ai/DeepSeek-V3.2 | generation-defaults | generation defaults changed (temperature 0.6 to 1); a call that does not pass these explicitly now generates differently (2025-12-01). |
| 2025-11-26 | allenai/Olmo-3-7B-Instruct | generation-defaults | generation defaults changed (do_sample unset to true); a call that does not pass these explicitly now generates differently (2025-11-26). |
| 2025-11-21 | allenai/Olmo-3-7B-Instruct | generation-defaults | generation defaults changed (max_new_tokens unset to 32768, temperature unset to 0.6, top_p unset to 0.95); a call that does not pass these explicitly now generates differently (2025-11-21). |
| 2025-11-21 | allenai/Olmo-3-7B-Think | generation-defaults | generation defaults changed (max_new_tokens unset to 32768, temperature unset to 0.6, top_p unset to 0.95); a call that does not pass these explicitly now generates differently (2025-11-21). |
| 2025-11-14 | swiss-ai/Apertus-8B-Instruct-2509 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-11-14). |
| 2025-11-07 | MiniMaxAI/MiniMax-M2 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-11-07). |
| 2025-11-06 | MiniMaxAI/MiniMax-M2 | structural tokenizer | EOS declared as 200020 (was unset); generation that relied on the library default now stops on this token; BOS declared as 200019 (was unset); model_type "minimax" to "minimax_m2"; a different class is instantiated from the same id (2025-11-06). |
| 2025-10-31 | allenai/Olmo-3-7B-Think | structural | architectures ["Olmo2RetrofitForCausalLM"] to ["Olmo3ForCausalLM"]; a different class is instantiated from the same id (2025-10-31). |
| 2025-10-27 | MiniMaxAI/MiniMax-M2 | structural | model_type "mixtral" to "minimax"; a different class is instantiated from the same id (2025-10-27). |
| 2025-10-27 | MiniMaxAI/MiniMax-M2 | generation-defaults | generation defaults changed (top_k unset to 40); a call that does not pass these explicitly now generates differently (2025-10-27). |
| 2025-10-25 | MiniMaxAI/MiniMax-M2 | structural | architectures ["MiniMaxM2MiniForCausalLM"] to ["MiniMaxM2ForCausalLM"]; a different class is instantiated from the same id (2025-10-25). |
| 2025-10-22 | moonshotai/Kimi-K2-Instruct | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-10-22). |
| 2025-10-22 | moonshotai/Kimi-K2-Instruct | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-10-22). |
| 2025-10-10 | moonshotai/Kimi-K2-Instruct | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-10-10). |
| 2025-10-08 | nvidia/Qwen2.5-VL-7B-Instruct-NVFP4 | structural | rope_scaling values changed ({"mrope_section":[16,24,24],"rope_type":"default","type":... to {"mrope_section":[16,24,24],"rope_type":"default","type":...); positions are computed differently (2025-10-08). |
| 2025-10-04 | amd/DeepSeek-R1-MXFP4 | structural | quantization_config changed; the stored format of the weights is different (2025-10-04). |
| 2025-09-23 | ai-sage/Giga-Embeddings-instruct | structural tokenizer | model_type "gigar" to "gigarembed"; a different class is instantiated from the same id; declared dtype "bfloat16" to "float32"; tokenizer_class "PreTrainedTokenizerFast" to "PreTrainedTokenizer" (2025-09-23). |
| 2025-09-23 | ai-sage/Giga-Embeddings-instruct | structural template | head dim 128 to 64, KV heads 2 to 16; derived parameters 3.30B to 2.50B (-801.0M, -1.60 GB of bf16 weights); rope_theta 1300 to 100000; every position embedding changes (2025-09-23). |
| 2025-09-22 | PrimeIntellect/Qwen3-0.6B | structural template | declared dtype "float32" to "bfloat16"; the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-09-22). |
| 2025-09-17 | Qwen/Qwen3-4B-Instruct-2507-FP8 | tokenizer template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering; tokenizer model_max_length 262144 to 1010000; truncation defaults change (2025-09-17). |
| 2025-09-17 | Qwen/Qwen3-4B-Instruct-2507 | tokenizer template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering; tokenizer model_max_length 262144 to 1010000; truncation defaults change (2025-09-17). |
| 2025-09-17 | Qwen/Qwen3-30B-A3B-Instruct-2507 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-09-17). |
| 2025-09-17 | Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 | tokenizer template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering; tokenizer model_max_length 262144 to 1010000; truncation defaults change (2025-09-17). |
| 2025-09-17 | Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 | tokenizer template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering; tokenizer model_max_length 262144 to 1010000; truncation defaults change (2025-09-17). |
| 2025-09-11 | OpenGVLab/InternVL3-8B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-09-11). |
| 2025-09-11 | OpenGVLab/InternVL3-1B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-09-11). |
| 2025-09-11 | OpenGVLab/InternVL2_5-4B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-09-11). |
| 2025-09-11 | OpenGVLab/InternVL2_5-4B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-09-11). |
| 2025-09-11 | OpenGVLab/InternVL3-1B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-09-11). |
| 2025-09-11 | OpenGVLab/InternVL3-8B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-09-11). |
| 2025-09-01 | swiss-ai/Apertus-8B-Instruct-2509 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-09-01). |
| 2025-09-01 | swiss-ai/Apertus-8B-Instruct-2509 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-09-01). |
| 2025-08-26 | deepseek-ai/DeepSeek-V3.1 | structural | quantization_config changed; the stored format of the weights is different (2025-08-26). |
| 2025-08-21 | nvidia/NVIDIA-Nemotron-Nano-9B-v2 | tokenizer | tokenizer model_max_length 1e+30 to 131072; truncation defaults change (2025-08-21). |
| 2025-08-21 | Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-08-21). |
| 2025-08-21 | Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-08-21). |
| 2025-08-21 | Qwen/Qwen3-Coder-30B-A3B-Instruct | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-08-21). |
| 2025-08-15 | nvidia/NVIDIA-Nemotron-Nano-9B-v2 | tokenizer template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering; add_bos_token true to false; encoding the same text yields different ids (2025-08-15). |
| 2025-08-13 | openai/gpt-oss-120b | tokenizer | EOS changed from [200002,199999] to [200002,199999,200012]; anything fine-tuned or served against the old value stops at a different token (2025-08-13). |
| 2025-08-13 | openai/gpt-oss-20b | tokenizer | EOS changed from [200002,199999] to [200002,199999,200012]; anything fine-tuned or served against the old value stops at a different token (2025-08-13). |
| 2025-08-11 | moonshotai/Kimi-K2-Instruct | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-08-11). |
| 2025-08-11 | moonshotai/Kimi-K2-Instruct | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-08-11). |
| 2025-08-08 | openai/gpt-oss-20b | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-08-08). |
| 2025-08-08 | openai/gpt-oss-120b | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-08-08). |
| 2025-08-08 | Qwen/Qwen3-30B-A3B-Instruct-2507 | tokenizer | tokenizer model_max_length 262144 to 1010000; truncation defaults change (2025-08-08). |
| 2025-08-07 | openai/gpt-oss-120b | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-08-07). |
| 2025-08-07 | openai/gpt-oss-20b | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-08-07). |
| 2025-08-07 | Qwen/Qwen3-Coder-30B-A3B-Instruct | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-08-07). |
| 2025-08-07 | Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-08-07). |
| 2025-08-07 | Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-08-07). |
| 2025-08-06 | unsloth/gpt-oss-20b-GGUF | structural | declared dtype unset to "bfloat16"; quantization_config changed; the stored format of the weights is different (2025-08-06). |
| 2025-08-05 | unsloth/gpt-oss-20b-BF16 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-08-05). |
| 2025-08-05 | openai/gpt-oss-120b | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-08-05). |
| 2025-08-05 | openai/gpt-oss-20b | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-08-05). |
| 2025-07-31 | Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8 | tokenizer generation-defaults | EOS changed from 151645 to [151645,151643]; anything fine-tuned or served against the old value stops at a different token; pad token unset to 151643; batched training masks a different id; generation defaults changed (do_sample unset to true, repetition_penalty unset to 1.05, temperature unset to 0.7, top_k unset to 20); a call that does not pass these explicitly now generates differently (2025-07-31). |
| 2025-07-29 | Qwen/Qwen3-30B-A3B-Instruct-2507 | generation-defaults | generation defaults changed (temperature 0.6 to 0.7, top_p 0.95 to 0.8); a call that does not pass these explicitly now generates differently (2025-07-29). |
| 2025-07-29 | Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 | tokenizer template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering; tokenizer model_max_length 131072 to 262144; truncation defaults change (2025-07-29). |
| 2025-07-29 | Qwen/Qwen3-30B-A3B-Instruct-2507 | tokenizer template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering; tokenizer model_max_length 131072 to 262144; truncation defaults change (2025-07-29). |
| 2025-07-29 | Qwen/Qwen3-30B-A3B-Instruct-2507 | generation-defaults | generation defaults changed (temperature 0.7 to 0.6, top_p 0.8 to 0.95); a call that does not pass these explicitly now generates differently (2025-07-29). |
| 2025-07-29 | Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 | tokenizer generation-defaults | EOS changed from 151645 to [151645,151643]; anything fine-tuned or served against the old value stops at a different token; pad token unset to 151643; batched training masks a different id; generation defaults changed (do_sample unset to true, temperature unset to 0.7, top_k unset to 20, top_p unset to 0.8); a call that does not pass these explicitly now generates differently (2025-07-29). |
| 2025-07-28 | HuggingFaceTB/SmolLM3-3B | structural | tie_word_embeddings unset to true; the output head is now shared with the embedding (2025-07-28). |
| 2025-07-28 | Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 | structural | MLP width 5504 to 6144; context window 1048576 to 262144 tokens; quantization_config changed; the stored format of the weights is different (2025-07-28). |
| 2025-07-28 | Qwen/Qwen3-30B-A3B-Instruct-2507 | structural | MLP width 5504 to 6144; context window 1048576 to 262144 tokens; architectures ["Qwen2MoeForCausalLM"] to ["Qwen3MoeForCausalLM"]; a different class is instantiated from the same id (2025-07-28). |
| 2025-07-22 | Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8 | tokenizer template generation-defaults | EOS changed from 151645 to [151645,151643]; anything fine-tuned or served against the old value stops at a different token; BOS 151643 removed; pad token unset to 151643; batched training masks a different id (2025-07-22). |
| 2025-07-22 | hmellor/Ilama-3.2-1B | structural | architectures ["LlamaForCausalLM"] to ["IlamaForCausalLM"]; a different class is instantiated from the same id (2025-07-22). |
| 2025-07-21 | Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 | tokenizer generation-defaults | EOS changed from 151645 to [151645,151643]; anything fine-tuned or served against the old value stops at a different token; pad token unset to 151643; batched training masks a different id; generation defaults changed (do_sample unset to true, temperature unset to 0.7, top_k unset to 20, top_p unset to 0.8); a call that does not pass these explicitly now generates differently (2025-07-21). |
| 2025-07-21 | Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 | tokenizer template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering; tokenizer model_max_length 131072 to 262144; truncation defaults change (2025-07-21). |
| 2025-07-21 | Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 | structural | MLP width 10944 to 12288; context window 1048576 to 262144 tokens; quantization_config changed; the stored format of the weights is different (2025-07-21). |
| 2025-07-18 | moonshotai/Kimi-K2-Instruct | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-07-18). |
| 2025-07-15 | moonshotai/Kimi-K2-Instruct | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-07-15). |
| 2025-07-15 | moonshotai/Kimi-K2-Instruct | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-07-15). |
| 2025-07-09 | HuggingFaceTB/SmolLM3-3B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-07-09). |
| 2025-07-08 | HuggingFaceTB/SmolLM3-3B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-07-08). |
| 2025-07-08 | HuggingFaceTB/SmolLM3-3B | generation-defaults | generation defaults changed (do_sample unset to true, temperature unset to 0.6, top_p unset to 0.95); a call that does not pass these explicitly now generates differently (2025-07-08). |
| 2025-07-08 | HuggingFaceTB/SmolLM3-3B | tokenizer | EOS changed from 128001 to 128012; anything fine-tuned or served against the old value stops at a different token (2025-07-08). |
| 2025-07-03 | jinaai/jina-embeddings-v4 | derived parameters 2.80B to 3.09B (+288.2M, +0.58 GB of bf16 weights); model_type declared as "qwen2_5_vl_text" (was unset) (2025-07-03). | |
| 2025-06-12 | amd/DeepSeek-R1-MXFP4 | tokenizer | tokenizer model_max_length 512 to 16384; truncation defaults change (2025-06-12). |
| 2025-06-04 | ai-sage/Giga-Embeddings-instruct | structural | context window 32768 to 4096 tokens (2025-06-04). |
| 2025-06-03 | Qwen/Qwen3-Embedding-0.6B | structural tokenizer generation-defaults | EOS changed from [151645,151643] to 151643; anything fine-tuned or served against the old value stops at a different token; context window 40960 to 32768 tokens; pad token 151643 to unset; batched training masks a different id (2025-06-03). |
| 2025-05-29 | deepseek-ai/DeepSeek-R1-0528-Qwen3-8B | structural | rope_scaling rewritten in a different format ({"factor":4,"original_max_position_embeddings":32768,"rop... to {"attn_factor":0.8782488562869419,"factor":4,"original_ma...); a loader that only reads the old format rejects or ignores the new one (2025-05-29). |
| 2025-05-29 | ai-sage/Giga-Embeddings-instruct | structural | layer count 27 to 36; derived parameters 2.61B to 3.30B (+693.7M, +1.39 GB of bf16 weights) (2025-05-29). |
| 2025-05-29 | deepseek-ai/DeepSeek-R1-0528 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-05-29). |
| 2025-05-20 | Qwen/Qwen3-8B-Base | tokenizer | EOS changed from "<|im_end|>" to "<|endoftext|>"; anything fine-tuned or served against the old value stops at a different token (2025-05-20). |
| 2025-05-20 | Qwen/Qwen3-4B-Base | tokenizer | EOS changed from "<|im_end|>" to "<|endoftext|>"; anything fine-tuned or served against the old value stops at a different token (2025-05-20). |
| 2025-05-20 | Qwen/Qwen3-1.7B-Base | tokenizer | EOS changed from "<|im_end|>" to "<|endoftext|>"; anything fine-tuned or served against the old value stops at a different token (2025-05-20). |
| 2025-05-20 | Qwen/Qwen3-0.6B-Base | tokenizer | EOS changed from "<|im_end|>" to "<|endoftext|>"; anything fine-tuned or served against the old value stops at a different token (2025-05-20). |
| 2025-05-19 | Qwen/Qwen3-8B-FP8 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-05-19). |
| 2025-05-19 | Qwen/Qwen3-8B-AWQ | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-05-19). |
| 2025-05-19 | Qwen/Qwen3-8B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-05-19). |
| 2025-05-19 | Qwen/Qwen3-4B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-05-19). |
| 2025-05-19 | Qwen/Qwen3-32B-AWQ | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-05-19). |
| 2025-05-19 | Qwen/Qwen3-32B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-05-19). |
| 2025-05-19 | Qwen/Qwen3-30B-A3B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-05-19). |
| 2025-05-19 | Qwen/Qwen3-235B-A22B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-05-19). |
| 2025-05-19 | Qwen/Qwen3-14B-AWQ | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-05-19). |
| 2025-05-19 | Qwen/Qwen3-14B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-05-19). |
| 2025-05-19 | Qwen/Qwen3-1.7B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-05-19). |
| 2025-05-19 | Qwen/Qwen3-0.6B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-05-19). |
| 2025-05-16 | jinaai/jina-embeddings-v4 | structural | derived parameters 3.09B to 2.80B (-288.3M, -0.58 GB of bf16 weights); architectures ["ColQwen25Duo"] to ["JinaEmbeddingsV4Model"]; a different class is instantiated from the same id (2025-05-16). |
| 2025-05-08 | Qwen/Qwen3-8B-AWQ | structural | declared dtype "bfloat16" to "float16" (2025-05-08). |
| 2025-05-06 | fixie-ai/ultravox-v0_5-llama-3_2-1b | tokenizer | tokenizer_class "PreTrainedTokenizerFast" to "PreTrainedTokenizer" (2025-05-06). |
| 2025-05-06 | DavidAU/Qwen3-8B-64k-Context-2X-Josiefied-Uncensored | structural | context window 40960 to 65536 tokens (2025-05-06). |
| 2025-05-06 | DavidAU/Qwen3-8B-64k-Context-2X-Josiefied-Uncensored | structural | rope_scaling added ({"factor":2,"original_max_position_embeddings":32768,"typ...); positions are now scaled (2025-05-06). |
| 2025-05-03 | Qwen/Qwen3-8B-AWQ | structural tokenizer template generation-defaults | EOS changed from 151645 to [151645,151643]; anything fine-tuned or served against the old value stops at a different token; BOS declared as 151643 (was unset); context window 65536 to 40960 tokens (2025-05-03). |
| 2025-05-01 | Qwen/Qwen3-32B-AWQ | structural tokenizer generation-defaults | EOS changed from 151645 to [151645,151643]; anything fine-tuned or served against the old value stops at a different token; BOS declared as 151643 (was unset); context window 65536 to 40960 tokens (2025-05-01). |
| 2025-05-01 | Qwen/Qwen3-14B-AWQ | structural tokenizer generation-defaults | EOS changed from 151645 to [151645,151643]; anything fine-tuned or served against the old value stops at a different token; BOS declared as 151643 (was unset); context window 65536 to 40960 tokens (2025-05-01). |
| 2025-04-28 | Qwen/Qwen3-8B-Base | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-04-28). |
| 2025-04-28 | Qwen/Qwen3-8B-FP8 | tokenizer generation-defaults | EOS changed from 151645 to [151645,151643]; anything fine-tuned or served against the old value stops at a different token; BOS declared as 151643 (was unset); generation defaults changed (do_sample unset to true, temperature unset to 0.6, top_k unset to 20, top_p unset to 0.95); a call that does not pass these explicitly now generates differently (2025-04-28). |
| 2025-04-28 | Qwen/Qwen3-4B-Base | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-04-28). |
| 2025-04-28 | Qwen/Qwen3-1.7B-Base | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-04-28). |
| 2025-04-28 | Qwen/Qwen3-0.6B-Base | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-04-28). |
| 2025-04-28 | Qwen/Qwen3-235B-A22B | structural | MLP width 8192 to 12288 (2025-04-28). |
| 2025-04-28 | Qwen/Qwen3-30B-A3B | structural | MLP width 8192 to 6144 (2025-04-28). |
| 2025-04-24 | ibm-granite/granite-3.3-2b-instruct | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-04-24). |
| 2025-04-24 | unsloth/Llama-3.2-3B-Instruct | tokenizer | tokenizer_class "PreTrainedTokenizerFast" to "PreTrainedTokenizer" (2025-04-24). |
| 2025-04-20 | moonshotai/Kimi-VL-A3B-Instruct | tokenizer | EOS declared as [163585] (was unset); generation that relied on the library default now stops on this token; BOS declared as 163584 (was unset); pad token unset to 163838; batched training masks a different id (2025-04-20). |
| 2025-04-20 | moonshotai/Kimi-VL-A3B-Instruct | generation-defaults | generation defaults changed (temperature 0.6 to 0.2); a call that does not pass these explicitly now generates differently (2025-04-20). |
| 2025-04-20 | moonshotai/Kimi-VL-A3B-Instruct | generation-defaults | generation defaults changed (do_sample unset to true, temperature 0.2 to 0.6); a call that does not pass these explicitly now generates differently (2025-04-20). |
| 2025-04-10 | ibm-granite/granite-3.3-2b-instruct | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-04-10). |
| 2025-04-10 | ibm-granite/granite-3.3-2b-instruct | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-04-10). |
| 2025-03-25 | deepseek-ai/DeepSeek-V3-0324 | tokenizer | tokenizer model_max_length 16384 to 131072; truncation defaults change (2025-03-25). |
| 2025-03-19 | google/gemma-3-1b-it | generation-defaults | generation defaults changed (do_sample unset to true, top_k unset to 64, top_p unset to 0.95); a call that does not pass these explicitly now generates differently (2025-03-19). |
| 2025-03-12 | google/gemma-3-1b-it | tokenizer | EOS changed from 1 to [1,106]; anything fine-tuned or served against the old value stops at a different token (2025-03-12). |
| 2025-03-12 | google/gemma-3-1b-it | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-03-12). |
| 2025-02-25 | unsloth/Llama-3.2-1B-Instruct | tokenizer | tokenizer_class "PreTrainedTokenizerFast" to "PreTrainedTokenizer" (2025-02-25). |
| 2025-02-18 | google/siglip2-giant-opt-patch16-384 | structural | architectures ["SiglipModel"] to unset; a different class is instantiated from the same id; declared dtype "float32" to unset (2025-02-18). |
| 2025-02-15 | unsloth/Meta-Llama-3.1-8B-Instruct | tokenizer | tokenizer_class "PreTrainedTokenizerFast" to "PreTrainedTokenizer" (2025-02-15). |
| 2025-02-09 | deepseek-ai/DeepSeek-R1-Distill-Qwen-7B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-02-09). |
| 2025-02-09 | deepseek-ai/DeepSeek-R1-Distill-Qwen-32B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-02-09). |
| 2025-02-09 | deepseek-ai/DeepSeek-R1-Distill-Qwen-14B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-02-09). |
| 2025-02-09 | deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-02-09). |
| 2025-02-09 | deepseek-ai/DeepSeek-R1 | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-02-09). |
| 2025-02-07 | fixie-ai/ultravox-v0_5-llama-3_2-1b | structural | layer count unset to 16; derived parameters 2.42B to 2.84B (+419.5M, +0.84 GB of bf16 weights) (2025-02-07). |
| 2025-02-07 | fixie-ai/ultravox-v0_5-llama-3_2-1b | structural | attention heads unset to 32; derived parameters 2.42B to 2.42B (+786.4K, +0.00 GB of bf16 weights) (2025-02-07). |
| 2025-02-04 | microsoft/phi-4 | tokenizer template | EOS changed from "<|endoftext|>" to "<|im_end|>"; anything fine-tuned or served against the old value stops at a different token; pad token "<|endoftext|>" to "<|dummy_85|>"; batched training masks a different id; the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-02-04). |
| 2025-02-02 | unsloth/Meta-Llama-3.1-8B-Instruct | structural tokenizer | EOS changed from [128001,128008,128009] to 128009; anything fine-tuned or served against the old value stops at a different token; head dim unset to 128; add_bos_token unset to true; encoding the same text yields different ids (2025-02-02). |
| 2025-01-23 | unsloth/Llama-3.2-3B-Instruct | tokenizer | EOS changed from [128001,128008,128009] to 128009; anything fine-tuned or served against the old value stops at a different token; add_bos_token unset to true; encoding the same text yields different ids (2025-01-23). |
| 2025-01-23 | unsloth/Llama-3.2-1B-Instruct | tokenizer | EOS changed from [128001,128008,128009] to 128009; anything fine-tuned or served against the old value stops at a different token (2025-01-23). |
| 2025-01-22 | unsloth/Llama-3.2-1B-Instruct | template | the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering (2025-01-22). |
| 2025-01-22 | deepseek-ai/DeepSeek-R1-Distill-Qwen-7B | tokenizer | add_bos_token false to true; encoding the same text yields different ids (2025-01-22). |
| 2025-01-22 | deepseek-ai/DeepSeek-R1-Distill-Qwen-32B | tokenizer | add_bos_token false to true; encoding the same text yields different ids (2025-01-22). |
| 2025-01-22 | deepseek-ai/DeepSeek-R1-Distill-Qwen-14B | tokenizer | add_bos_token false to true; encoding the same text yields different ids (2025-01-22). |
| 2025-01-22 | deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B | tokenizer | add_bos_token false to true; encoding the same text yields different ids (2025-01-22). |
| 2025-01-22 | deepseek-ai/DeepSeek-R1 | tokenizer | add_bos_token false to true; encoding the same text yields different ids (2025-01-22). |
| 2025-01-22 | Alibaba-NLP/gte-reranker-modernbert-base | tokenizer | tokenizer model_max_length 1e+30 to 8192; truncation defaults change (2025-01-22). |
| 2025-01-22 | unsloth/Llama-3.2-1B-Instruct | tokenizer template generation-defaults | pad token unset to 128004; batched training masks a different id; the chat template changed; prompts rendered with the old one no longer match what the model now expects, so cached prompts and eval sets need re-rendering; add_bos_token unset to true; encoding the same text yields different ids (2025-01-22). |
| 2025-01-22 | deepseek-ai/DeepSeek-R1-Distill-Qwen-7B | tokenizer | add_bos_token true to false; encoding the same text yields different ids (2025-01-22). |
| 2025-01-22 | deepseek-ai/DeepSeek-R1-Distill-Qwen-32B | tokenizer | add_bos_token true to false; encoding the same text yields different ids (2025-01-22). |
| 2025-01-22 | deepseek-ai/DeepSeek-R1-Distill-Qwen-14B | tokenizer | add_bos_token true to false; encoding the same text yields different ids (2025-01-22). |
| 2025-01-22 | deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B | tokenizer | add_bos_token true to false; encoding the same text yields different ids (2025-01-22). |
| 2025-01-22 | deepseek-ai/DeepSeek-R1 | tokenizer | add_bos_token true to false; encoding the same text yields different ids (2025-01-22). |
| 2025-01-16 | unsloth/Llama-3.2-3B-Instruct | tokenizer generation-defaults | pad token unset to 128004; batched training masks a different id; generation defaults changed (max_length unset to 131072); a call that does not pass these explicitly now generates differently (2025-01-16). |
| 2025-01-15 | answerdotai/ModernBERT-base | tokenizer | tokenizer model_max_length 1e+30 to 8192; truncation defaults change (2025-01-15). |
| 2025-01-11 | Alibaba-NLP/gte-Qwen2-7B-instruct | structural | config.json was removed; whatever defaults it set now come from the library (2025-01-11). |
| 2025-01-07 | HuggingFaceTB/SmolLM2-135M-Instruct | tokenizer | tokenizer model_max_length 2048 to 8192; truncation defaults change (2025-01-07). |
| 2025-01-07 | HuggingFaceTB/SmolLM2-360M-Instruct | tokenizer | tokenizer model_max_length 2048 to 8192; truncation defaults change (2025-01-07). |
| 2025-01-07 | HuggingFaceTB/SmolLM2-1.7B-Instruct | tokenizer | tokenizer model_max_length 2048 to 8192; truncation defaults change (2025-01-07). |
| 2025-01-07 | HuggingFaceTB/SmolLM2-135M-Instruct | tokenizer | tokenizer model_max_length 8192 to 2048; truncation defaults change (2025-01-07). |
| 2025-01-07 | HuggingFaceTB/SmolLM2-135M-Instruct | tokenizer | tokenizer model_max_length 2048 to 8192; truncation defaults change (2025-01-07). |
| 2024-12-19 | answerdotai/ModernBERT-base | structural | derived parameters 130.1M to 136.0M (+5.9M, +0.01 GB of bf16 weights); context window 512 to 8192 tokens (2024-12-19). |
| 2024-12-19 | ai-sage/Giga-Embeddings-instruct | tokenizer | pad token unset to "<unk>"; batched training masks a different id (2024-12-19). |
| 2024-12-16 | utter-project/EuroLLM-1.7B-Instruct | tokenizer | EOS changed from 2 to 4; anything fine-tuned or served against the old value stops at a different token; BOS changed from 1 to 3; prompts built by the old tokenizer start with a different token (2024-12-16). |
| 2024-12-11 | answerdotai/ModernBERT-base | structural | architectures ["ModernBertModel"] to ["ModernBertForMaskedLM"]; a different class is instantiated from the same id (2024-12-11). |
| 2024-12-08 | microsoft/Florence-2-large | structural | derived parameters 204.7M to 207.8M (+3.1M, +0.01 GB of bf16 weights); context window 1024 to 4096 tokens (2024-12-08). |
| 2024-12-06 | meta-llama/Llama-3.1-405B-FP8 | tokenizer | pad token unset to "<|finetune_right_pad_id|>"; batched training masks a different id (2024-12-06). |
| 2024-12-06 | OpenGVLab/InternVL2-1B | structural | derived parameters 798.6M to 934.6M (+136.0M, +0.27 GB of bf16 weights); tie_word_embeddings true to false; the output head is no longer shared with the embedding (2024-12-06). |
| 2024-11-29 | OpenGVLab/InternVL2_5-4B | structural | derived parameters 3.39B to 3.70B (+310.8M, +0.62 GB of bf16 weights); tie_word_embeddings true to false; the output head is no longer shared with the embedding (2024-11-29). |
| 2024-11-26 | OpenGVLab/InternVL2_5-4B | tokenizer | EOS changed from [151644,151645] to [151644,151645,151643]; anything fine-tuned or served against the old value stops at a different token (2024-11-26). |
| 2024-11-22 | jinaai/jina-clip-v2 | structural | declared dtype "float16" to "bfloat16" (2024-11-22). |
| 2024-11-22 | jinaai/jina-clip-v2 | structural | declared dtype "bfloat16" to "float16" (2024-11-22). |
| 2024-11-22 | jinaai/jina-clip-v2 | declared dtype unset to "bfloat16" (2024-11-22). | |
| 2024-11-18 | Snowflake/snowflake-arctic-embed-m-v2.0 | structural | architectures ["NewModel"] to ["GteModel"]; a different class is instantiated from the same id (2024-11-18). |
| 2024-11-18 | Qwen/Qwen2.5-Coder-14B-Instruct-AWQ | tokenizer | tokenizer model_max_length 131072 to 32768; truncation defaults change (2024-11-18). |
| 2024-11-18 | Qwen/Qwen2.5-Coder-14B-Instruct | tokenizer | tokenizer model_max_length 131072 to 32768; truncation defaults change (2024-11-18). |
| 2024-11-18 | Qwen/Qwen2.5-Coder-32B-Instruct-AWQ | tokenizer | tokenizer model_max_length 131072 to 32768; truncation defaults change (2024-11-18). |
| 2024-11-18 | Qwen/Qwen2.5-Coder-32B-Instruct | tokenizer | tokenizer model_max_length 131072 to 32768; truncation defaults change (2024-11-18). |
| 2024-11-18 | Qwen/Qwen2.5-Coder-7B-Instruct-AWQ | tokenizer | EOS changed from 151643 to 151645; anything fine-tuned or served against the old value stops at a different token; tokenizer model_max_length 131072 to 32768; truncation defaults change (2024-11-18). |
| 2024-11-18 | Qwen/Qwen2.5-Coder-7B-Instruct-GPTQ-Int4 | tokenizer | EOS changed from 151643 to 151645; anything fine-tuned or served against the old value stops at a different token; tokenizer model_max_length 131072 to 32768; truncation defaults change (2024-11-18). |
| 2024-11-18 | Qwen/Qwen2.5-Coder-7B-Instruct | tokenizer | tokenizer model_max_length 131072 to 32768; truncation defaults change (2024-11-18). |
| 2024-11-18 | Qwen/Qwen2.5-Coder-7B | tokenizer | tokenizer model_max_length 131072 to 32768; truncation defaults change (2024-11-18). |
| 2024-11-18 | Qwen/Qwen2.5-Coder-3B-Instruct | tokenizer | tokenizer model_max_length 131072 to 32768; truncation defaults change (2024-11-18). |
| 2024-11-18 | Qwen/Qwen2.5-Coder-1.5B-Instruct | tokenizer | tokenizer model_max_length 131072 to 32768; truncation defaults change (2024-11-18). |
| 2024-11-18 | Qwen/Qwen2.5-Coder-1.5B | tokenizer | tokenizer model_max_length 131072 to 32768; truncation defaults change (2024-11-18). |
| 2024-11-15 | Alibaba-NLP/gte-Qwen2-1.5B-instruct | structural tokenizer | context window 2048 to 131072 tokens; sliding_window 2048 to 131072; tokenizer model_max_length 2048 to 32768; truncation defaults change (2024-11-15). |
| 2024-11-12 | jinaai/jina-clip-v2 | structural | tie_word_embeddings true removed; the library default decides whether the head is shared; declared dtype "float16" to unset (2024-11-12). |
| 2024-11-12 | Alibaba-NLP/gte-Qwen2-1.5B-instruct | structural tokenizer | context window 131072 to 2048 tokens; sliding_window 131072 to 2048; tokenizer model_max_length 32768 to 2048; truncation defaults change (2024-11-12). |
| 2024-11-12 | Qwen/Qwen2.5-Coder-1.5B-Instruct | structural | sliding_window 131072 to 32768 (2024-11-12). |
| 2024-11-12 | Qwen/Qwen2.5-Coder-1.5B | structural | sliding_window 131072 to 32768 (2024-11-12). |
| 2024-11-09 | Qwen/Qwen2.5-Coder-32B-Instruct | tokenizer | EOS changed from 151643 to 151645; anything fine-tuned or served against the old value stops at a different token (2024-11-09). |
| 2024-11-09 | Qwen/Qwen2.5-Coder-32B-Instruct | generation-defaults | generation defaults changed (repetition_penalty 1.1 to 1.05); a call that does not pass these explicitly now generates differently (2024-11-09). |
| 2024-11-09 | Qwen/Qwen2.5-Coder-14B-Instruct | tokenizer | EOS changed from 151643 to 151645; anything fine-tuned or served against the old value stops at a different token (2024-11-09). |
| 2024-11-09 | Qwen/Qwen2.5-Coder-14B-Instruct | generation-defaults | generation defaults changed (repetition_penalty 1.1 to 1.05); a call that does not pass these explicitly now generates differently (2024-11-09). |
| 2024-11-09 | Qwen/Qwen2.5-Coder-7B-Instruct | tokenizer | EOS changed from 151643 to 151645; anything fine-tuned or served against the old value stops at a different token (2024-11-09). |
Showing the newest 300 of 586. Every row as JSON.
How this is read
Every day the tracker asks the Hugging Face Hub for the head commit of each tracked model. Only when that commit moved does it read config.json, generation_config.json, tokenizer_config.json and chat_template.jinja at the new commit, keep a normalised subset of the keys that change what the model computes, how it tokenises, or how it generates by default, and diff it against the last snapshot.
History before the first daily run comes from walking each repository's git history for those four files, which is complete for them and blind to the weights: a weight swap with no file change is only visible from the first daily snapshot on.
Chat templates are stored as a sha256 prefix, never in full. Aliases are folded (torch_dtype and dtype, rope_theta inside rope_parameters) so a re-save under a newer transformers is not reported as a change.
Parameter and memory deltas come from reconstructing the graph from each config with the same code the /m/ pages use. Every consequence sentence says what the key does mechanically; none is a report that something broke for anyone.
Pin what you tested
from_pretrained("zai-org/GLM-5.3", revision="ba32f8c19c8738e2c57c71798c601d477a7bf844")