Recent Open Weights Model Launches

Artificial Analysis reports that the three highest-scoring open weights models, whose trained weights anyone can download and run, were all released in the week before this April 30 roundup (artificialanalysis.ai). Moonshot AI's Kimi K2.6 and Xiaomi's MiMo V2.5 Pro tie at 54 on its Intelligence Index, with DeepSeek's V4 Pro at 52. The leading proprietary models are OpenAI's GPT-5.5 at 60, then Google's Gemini 3.1 Pro Preview and Anthropic's Claude Opus 4.7 at 57. That leaves the best open models 3 to 6 points behind. A year earlier the top open model, DeepSeek V3 0324, scored 22, about 13 points below Claude 3.7 Sonnet at 35.

The single index number hides an uneven picture. On HLE (Humanity's Last Exam) the three open models score 34 to 36%, against 44% for GPT-5.5 and 45% for Gemini 3.1 Pro Preview. On CritPt, a research-level physics test, they score 4 to 12% against 27% for GPT-5.5, and on TerminalBench Hard, which tests agentic coding and terminal use, 43 to 46% against 61% for GPT-5.5 and 54% for Gemini. AA-Omniscience, which measures knowledge while penalizing hallucinated answers, shows a large gap too: DeepSeek V4 Pro scores -10, MiMo V2.5 Pro +4 and Kimi K2.6 +6, while GPT-5.5 scores +20, Opus 4.7 +26 and Gemini 3.1 Pro Preview +33.

Where the open models win is price. All three are mixture-of-experts models with more than a trillion parameters, a design that uses only part of the model for each token (Kimi K2.6 activates 32B of its 1T), and all carry permissive licenses. Nine of the 13 models on Artificial Analysis's intelligence-versus-price Pareto frontier, the set of models that no other model beats on both score and price, are open weights, and it says the top three offer intelligence comparable to leading proprietary models at between half and one-sixth of the price. All of the top 10 open weights models come from China-based labs; the highest-ranked from elsewhere are Gemma 4 31B and NVIDIA Nemotron 3 Super.