Model Database

Pre-computed PeakWeights analysis for top open-source LLMs in 2026.

Analyzed Models

Model	Params	FP16	4-bit	Protected	Recovery	K	Type	Download
Qwen2.5-7B Alibaba	7.6B	10.62	11.51	10.63	99%	50	MLP	.pwi
Mistral-7B-v0.3 Mistral AI	7.2B	13.44	13.8	13.58	61%	100	lm_head	.pwi
SmolLM2-1.7B HuggingFace	1.7B	17.92	24.56	17.99	99%	50	MLP	.pwi
DeepSeek-R1-Distill-Qwen-7B DeepSeek	7.6B	70.73	73.52	70.91	93%	50	MLP	.pwi
Phi-3-mini Microsoft	3.8B	11.65	12.67	11.68	97%	50	MLP	.pwi

Perplexity measured on WikiText-103. Recovery at K=50 protected weights.

Top 25 Models (Analysis Pending)

Run locally with:peakweights MODEL_ID --output weights.pwi

Model	Organization	Parameters	Architecture	License	Link
DeepSeek-V3	DeepSeek	671B (37B active)	MoE	MIT	View
DeepSeek-R1	DeepSeek	671B (37B active)	MoE	MIT	View
Qwen3-235B-A22B	Alibaba	235B (22B active)	MoE	Qwen License	View
Llama 4 Maverick	Meta	400B (17B active)	MoE	Llama 4 License	View
Llama 4 Scout	Meta	109B (17B active)	MoE	Llama 4 License	View
Qwen2.5-72B-Instruct	Alibaba	72B	Dense Transformer	Qwen License	View
Llama-3.3-70B-Instruct	Meta	70B	Dense Transformer	Llama 3.3 License	View
DeepSeek-R1-Distill-Llama-70B	DeepSeek	70B	Dense Transformer	MIT	View
Mixtral-8x22B-Instruct-v0.1	Mistral AI	141B (39B active)	MoE	Apache 2.0	View
Qwen2.5-14B-Instruct	Alibaba	14B	Dense Transformer	Apache 2.0	View
Qwen2.5-7B-Instruct	Alibaba	7B	Dense Transformer	Apache 2.0	View
DeepSeek-R1-Distill-Qwen-7B	DeepSeek	7B	Dense Transformer	MIT	View
Mistral-7B-Instruct-v0.3	Mistral AI	7B	Dense Transformer	Apache 2.0	View
Gemma-2-9B-it	Google	9B	Dense Transformer	Gemma License	View
Gemma-2-27B-it	Google	27B	Dense Transformer	Gemma License	View
Phi-4	Microsoft	14B	Dense Transformer	MIT	View
Qwen2.5-Coder-32B-Instruct	Alibaba	32B	Dense Transformer	Apache 2.0	View
Codestral-22B-v0.1	Mistral AI	22B	Dense Transformer	MNPL	View
DeepSeek-Coder-V2-Instruct	DeepSeek	236B (21B active)	MoE	MIT	View
SmolLM2-1.7B-Instruct	HuggingFace	1.7B	Dense Transformer	Apache 2.0	View
Qwen2.5-3B-Instruct	Alibaba	3B	Dense Transformer	Apache 2.0	View
Phi-3.5-mini-instruct	Microsoft	3.8B	Dense Transformer	MIT	View
Gemma-2-2B-it	Google	2B	Dense Transformer	Gemma License	View
MiMo-V2-Flash	Xiaomi	309B (15B active)	MoE	Apache 2.0	View
Command R+	Cohere	104B	Dense Transformer	CC-BY-NC	View
OLMo-2-13B-Instruct	AI2	13B	Dense Transformer	Apache 2.0	View

Run your own analysis

Analyze any HuggingFace model with the CLI or Python API.

CLI

pip install peakweights

# Analyze model
peakweights Qwen/Qwen2.5-7B \
  --top_k 50 \
  --output weights.pwi

Python

from peakweights import find, save

critical = find(
  "Qwen/Qwen2.5-7B",
  k=50,
  device="cuda"
)

save(critical, "weights.pwi")