🔓 Open Source

Open-Weights LLM Leaderboard

Rankings of open-source and open-weights language models. 9 models from 4 providers.

#	Model	Composite	ELO	MMLU	HumanEval	Speed	Params
1	DeepSeek R1 DeepSeek Open	92.3	1358	90.8%	92.1%	32 t/s	671B (37B active MoE)
2	Llama 4 Maverick Meta Open	87.5	1325	89.5%	90.8%	120 t/s	400B (17B active MoE)
3	DeepSeek V3 DeepSeek Open	86.1	1318	88.5%	89.4%	68 t/s	671B (37B active MoE)
4	Qwen 3 235B Alibaba Open	84.5	1305	87.8%	88.7%	95 t/s	235B (22B active MoE)
5	Llama 3.1 405B Meta Open	82.1	1265	87.3%	89%	29 t/s	405B
6	Llama 3.3 70B Meta Open	79.8	1258	86%	86.4%	65 t/s	70B
7	Qwen 2.5 72B Alibaba Open	76	1230	85.2%	82.4%	55 t/s	72B
8	Llama 3.1 70B Meta Open	74.3	1235	83.7%	80.5%	62 t/s	70B
9	Mixtral 8x22B Mistral AI Open	65.2	1195	77.8%	75.8%	82 t/s	176B (39B active MoE)