Architectures / Real checkpoints
We derive a model's size from its structure. Here is every time we checked.
15 of 28 comparable checkpoints are reproduced within 2% from config alone; the median miss is 1.3%.
Each row reconstructs a checkpoint from its config.json with no weights read, then puts our
parameter count next to the number of scalars actually in the published weight files.
Excluded from that rate, by a rule read off the config before any delta was computed: 4 quantized, 8 multi-tower, 0 weight-sharing. They are still in the table, with the reason on each page.
| Model | Downloads | Derived | Published | Delta | Note |
|---|---|---|---|---|---|
| all-MiniLM-L6-v2 | 254M | 22.6M | 22.7M | -0.68% | |
| all-mpnet-base-v2 | 24.5M | 109M | 109M | -0.56% | |
| t5-small | 23.4M | 60.6M | 60.5M | +0.11% | |
| Qwen3-0.6B | 21.9M | 596M | 752M | -20.7% | |
| nomic-embed-text-v1.5 | 16.5M | 110M | 137M | -19.5% | |
| gpt2 | 14.6M | 124M | 137M | -9.18% | |
| Qwen2.5-7B-Instruct | 11.5M | 7.62B | 7.62B | +0.00% | |
| Qwen3.6-35B-A3B-NVFP4 | 10.3M | 34.01B | 18.68B | +82.0% | quantized |
| roberta-base | 9.7M | 124M | 125M | -0.53% | |
| clap-htsat-fused | 9.2M | 152M | 154M | -0.77% | multi-tower |
| distilbert-base-uncased | 7.1M | 66.4M | 67.0M | -0.95% | |
| whisper-large-v3-turbo | 7.1M | 801M | 809M | -0.91% | |
| OTel-2.0-LLM-31B-IT | 6.8M | 30.56B | 31.27B | -2.26% | multi-tower |
| gpt-oss-20b | 6.4M | 20.91B | 20.91B | -0.03% | quantized |
| granite-embedding-small-english-r2 | 6.1M | 43.8M | 47.7M | -8.16% | |
| vit-base-patch16-224 | 5.1M | 85.4M | 86.6M | -1.40% | |
| dolphin-2.9.1-yi-1.5-34b | 4.8M | 34.39B | 34.39B | +0.00% | |
| DeepSeek-V4-Flash-0731 | 4.6M | 291B | 304B | -4.35% | quantized |
| dinov2-small | 4.4M | 47.4M | 22.1M | +115% | |
| Qwen3-ASR-1.7B | 4.0M | 152M | 2.35B | -93.5% | multi-tower |
| Qwen-72B | 3.7M | 121B | 72.29B | +66.8% | |
| pythia-160m | 3.4M | 164M | 213M | -22.9% | |
| bart-large-mnli | 3.2M | 405M | 407M | -0.53% | |
| Qwen3-TTS-12Hz-1.7B-CustomVoice | 2.6M | 152M | 1.92B | -92.1% | multi-tower |
| koelectra-small-v3-nsmc | 2.5M | 18.6M | 14.1M | +31.4% | |
| siglip2-giant-opt-patch16-384 | 2.5M | 1.84B | 1.87B | -1.46% | multi-tower |
| w2v-bert-2.0 | 2.3M | 739M | 580M | +27.4% | |
| jina-embeddings-v3 | 2.3M | 567M | 572M | -0.99% | |
| Qwen3-30B-A3B | 2.3M | 30.53B | 30.53B | +0.00% | |
| Voxtral-Mini-4B-Realtime-2602 | 2.2M | 4.06B | 4.43B | -8.28% | multi-tower |
| ESMC-6B | 2.1M | 3.36B | 6.35B | -47.1% | |
| resnet-50 | 1.6M | 25.5M | 25.6M | -0.32% | |
| esm2_t33_650M_UR50D | 1.4M | 651M | 652M | -0.27% | |
| Qwen3-VL-Embedding-2B | 1.4M | 1.72B | 2.13B | -19.0% | multi-tower |
| gte-multilingual-base | 1.3M | 283M | 305M | -7.21% | |
| clipseg-rd64-refined | 1.3M | 85.4M | 151M | -43.4% | multi-tower |
| table-transformer-structure-recognition | 1.2M | 16.0M | 28.8M | -44.7% | |
| camembert-base | 888K | 110M | 111M | -1.11% | |
| rtdetr_r101vd_coco_o365 | 765K | 9.3M | 76.8M | -87.9% | |
| e5-mistral-7b-instruct-bnb-4bit | 763K | 7.24B | 7.33B | -1.19% | quantized |
Why publish the misses
Because they are the only reason to believe the hits. A derived number with nothing to check it against is a number you have to take our word for, and this table exists so nobody has to.
It has already paid for itself once: diffing these rows is what found that distilbert-base-uncased came back as BERT-base, 62% too large, because four of its config keys were missing from our alias table. That was live, silent, and nothing in the product would have told you.
Held back (15)
Candidates from the same download-ordered pool that never became a page. Listed because a set of 40 assembled from a larger pool is a selection.
- BAAI/bge-m3: no safetensors index on the Hub, so there is nothing to check our count against
- jonatasgrosman/wav2vec2-large-xlsr-53-japanese: no safetensors index on the Hub, so there is nothing to check our count against
- openai/clip-vit-base-patch32: no safetensors index on the Hub, so there is nothing to check our count against
- microsoft/mdeberta-v3-base: no safetensors index on the Hub, so there is nothing to check our count against
- stabilityai/stable-diffusion-xl-base-1.0: config.json is not readable (gated, private or absent)
- facebook/opt-125m: no safetensors index on the Hub, so there is nothing to check our count against
- Helsinki-NLP/opus-mt-nl-en: no safetensors index on the Hub, so there is nothing to check our count against
- depth-anything/Depth-Anything-V2-Small-hf: our own verifier says this graph will not forward-pass (Empty graph)
- kha-white/manga-ocr-base: no safetensors index on the Hub, so there is nothing to check our count against
- Salesforce/blip-image-captioning-base: no safetensors index on the Hub, so there is nothing to check our count against
- microsoft/wavlm-large: no safetensors index on the Hub, so there is nothing to check our count against
- facebook/nllb-200-distilled-600M: no safetensors index on the Hub, so there is nothing to check our count against
- nvidia/segformer-b0-finetuned-ade-512-512: our own verifier says this graph will not forward-pass (Empty graph)
- google/siglip2-base-patch16-naflex: our own verifier says this graph will not forward-pass (Empty graph)
- Qwen/Qwen-Image-Edit-2509: config.json is not readable (gated, private or absent)
Do this to your own model
pip install neurarch-trace
neurarch-trace path/to/model.py:MyModel --input 1,3,224,224 --plan --share