HumanEval、MBPP 和 SWE-bench 综合。

编程最佳 AI 模型.

按已发布的编程基准排名。SWE-bench(开源仓库中的真实 bug)权重最高 — 最能预测 agent 行为。HumanEval(函数合成)和 MBPP(小型 Python 程序)填底层能力。

使用的基准: HUMANEVAL · 30% MBPP · 20% SWE BENCH · 50%

#	模型	分数	来自
1	Claude 3.5 Sonnet closed	92.0	Anthropic
2	Mistral Large 2 open	92.0	Mistral AI
3	Qwen 2.5 Coder 32B open	91.7	Alibaba (Qwen Team)
4	GPT-4o closed	90.2	OpenAI
5	DeepSeek R1 open	90.0	DeepSeek
6	Llama 3.1 405B open	89.0	Meta AI
7	Grok 3 closed	88.4	xAI
8	Llama 3.3 70B open	88.4	Meta AI
9	GPT-4o Mini closed	87.2	OpenAI
10	Qwen 2.5 72B open	86.6	Alibaba (Qwen Team)
11	Claude Haiku 4.5 closed	83.0	Anthropic
12	DeepSeek R1 Distill Llama 70B open	83.0	DeepSeek
13	DeepSeek V3 open	82.6	DeepSeek
14	Claude Sonnet 4.6 closed	82.6	Anthropic
15	GPT-5 closed	81.8	OpenAI
16	Gemma 3 27B open	81.0	Google DeepMind
17	Llama 3.1 70B open	80.5	Meta AI
18	DeepSeek R1 Distill Qwen 32B open	80.0	DeepSeek
19	Claude Opus 4.7 closed	79.8	Anthropic
20	Mixtral 8x22B open	76.0	Mistral AI
21	Gemini 2.5 Pro closed	72.7	Google DeepMind
22	Llama 3.1 8B open	72.6	Meta AI
23	Command R+ open	70.7	Cohere
24	Kimi K2 open	65.8	Moonshot AI

Showing top 24 models with published data on at least one of the benchmarks above. Scores are weighted averages on a 0–100 scale.

AI 模型排行榜

More leaderboards.