The best open-weights models sit 3 to 6 index points behind proprietary leaders and dominate the price frontier, but the gap remains wide on hard reasoning and agentic coding.
Key takeaways · AI-distilled
Kimi K2.6 and MiMo V2.5 Pro tie at 54 on the Intelligence Index with DeepSeek V4 Pro at 52, within 3 to 6 points of GPT-5.5 (xhigh) at 60. A year earlier, the top open model scored 22, about 13 points behind the proprietary leader.
The gap stays wide on the hardest evalA repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.Full definition →: the top open models score 34-36% on HLE versus 44-45% for GPT-5.5 and Gemini 3.1 Pro Preview, and 43-46% on TerminalBench Hard versus 61% for GPT-5.5 (xhigh).
On AA-Omniscience, DeepSeek V4 Pro scores -10, hallucinating more than MiMo V2.5 Pro (+4) and Kimi K2.6 (+6), while GPT-5.5, Claude Opus 4.7 and Gemini 3.1 Pro Preview score +20, +26 and +33.
All top 10 open-weights models come from China-based labs, and 9 of the 13 models on the intelligence-versus-price Pareto frontier are open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition →, with comparable intelligence at one-half to one-sixth of the price.
Terms in this piece · Glossary
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
