N Neurarch Architectures Models Checks Data Docs Open the app

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.

Checkpoints
40
Comparable
28
Within 2%
15
Median miss
1.3%

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.

ModelDownloadsDerivedPublishedDeltaNote
all-MiniLM-L6-v2254M22.6M22.7M-0.68%
all-mpnet-base-v224.5M109M109M-0.56%
t5-small23.4M60.6M60.5M+0.11%
Qwen3-0.6B21.9M596M752M-20.7%
nomic-embed-text-v1.516.5M110M137M-19.5%
gpt214.6M124M137M-9.18%
Qwen2.5-7B-Instruct11.5M7.62B7.62B+0.00%
Qwen3.6-35B-A3B-NVFP410.3M34.01B18.68B+82.0%quantized
roberta-base9.7M124M125M-0.53%
clap-htsat-fused9.2M152M154M-0.77%multi-tower
distilbert-base-uncased7.1M66.4M67.0M-0.95%
whisper-large-v3-turbo7.1M801M809M-0.91%
OTel-2.0-LLM-31B-IT6.8M30.56B31.27B-2.26%multi-tower
gpt-oss-20b6.4M20.91B20.91B-0.03%quantized
granite-embedding-small-english-r26.1M43.8M47.7M-8.16%
vit-base-patch16-2245.1M85.4M86.6M-1.40%
dolphin-2.9.1-yi-1.5-34b4.8M34.39B34.39B+0.00%
DeepSeek-V4-Flash-07314.6M291B304B-4.35%quantized
dinov2-small4.4M47.4M22.1M+115%
Qwen3-ASR-1.7B4.0M152M2.35B-93.5%multi-tower
Qwen-72B3.7M121B72.29B+66.8%
pythia-160m3.4M164M213M-22.9%
bart-large-mnli3.2M405M407M-0.53%
Qwen3-TTS-12Hz-1.7B-CustomVoice2.6M152M1.92B-92.1%multi-tower
koelectra-small-v3-nsmc2.5M18.6M14.1M+31.4%
siglip2-giant-opt-patch16-3842.5M1.84B1.87B-1.46%multi-tower
w2v-bert-2.02.3M739M580M+27.4%
jina-embeddings-v32.3M567M572M-0.99%
Qwen3-30B-A3B2.3M30.53B30.53B+0.00%
Voxtral-Mini-4B-Realtime-26022.2M4.06B4.43B-8.28%multi-tower
ESMC-6B2.1M3.36B6.35B-47.1%
resnet-501.6M25.5M25.6M-0.32%
esm2_t33_650M_UR50D1.4M651M652M-0.27%
Qwen3-VL-Embedding-2B1.4M1.72B2.13B-19.0%multi-tower
gte-multilingual-base1.3M283M305M-7.21%
clipseg-rd64-refined1.3M85.4M151M-43.4%multi-tower
table-transformer-structure-recognition1.2M16.0M28.8M-44.7%
camembert-base888K110M111M-1.11%
rtdetr_r101vd_coco_o365765K9.3M76.8M-87.9%
e5-mistral-7b-instruct-bnb-4bit763K7.24B7.33B-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.

Do this to your own model

pip install neurarch-trace
neurarch-trace path/to/model.py:MyModel --input 1,3,224,224 --plan --share