AI
Kimi K3.1 Pro opens model weights
A technical article reposted on Hacker News says Kimi K3.1 Pro is an open-weight, natively multimodal agent model in the 3-trillion-parameter class, supporting a 1-million-token context. The article says it was post-trained from Kimi K3 through large-scale joint reinforcement learning, with notable improvements in coding, 3D creation, financial research, and factual accuracy.
The article says Kimi K3.1 Pro is the first model in the K3.1 series and Moonshot AI's most capable 3T-class open-weight model to date. The model natively supports multimodal and agent tasks, with a context window of 1 million tokens.
According to the article, Kimi K3.1 Pro continues post-training from Kimi K3 and underwent large-scale joint reinforcement learning across hundreds of downstream domains. Improvements include coding, 3D creation, financial research, factual accuracy, and other agent capabilities.
Architecturally, the article says Kimi K3.1 Pro replaces the full-attention Gated MLA layers in Kimi K3 with Gated DeepSeek Sparse Attention layers to improve long-context modeling efficiency. Unless otherwise specified, all results in the article used maximum reasoning effort, a temperature of 1.0, and top-p of 0.95.
The material also lists multiple evaluation footnotes involving benchmarks such as Terminal-Bench, SWE-Marathon, ProgramBench, DeepSWE, and BenchCAD, and explains that different tasks used tools such as Claude Code, Kimi Code, or Codex, with some scores coming from official leaderboards, model blogs, or internal evaluations. The article cautions that Elo scores and benchmark-specific metrics cannot be directly compared with accuracy or pass rates.