Moonshot releases Kimi K3.1 Pro and promises open weights by November 4
Moonshot AI's new model has a 1M-token context and, by the company's own measurements, nearly matches Claude Opus 5.5 on terminal coding tasks. The weights are not public yet.
Source: Moonshot AI (блог Kimi) · Authors: Moonshot AI · License: Цитата из официального поста
Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence
We’re introducing Kimi K3.1 Pro, the first model in our new K3.1 series and our most capable 3T-class open-weight model to date, with a 1-million-token context window. It delivers substantial improvements across coding, 3D creation, financial research, and other agentic capabilities. We are also releasing Kimi K3.1 Pro Highspeed, which delivers 2-3× the speed of K3.1 Pro for more responsive interactions. Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed are available today on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. The full model weights will be released by Nov 4.
This is an excerpt (abstract or key part). The full text is at the source link.
What shipped. Moonshot AI released Kimi K3.1 Pro and a faster K3.1 Pro Highspeed variant (2–3× faster, per the company). It is a post-trained version of Kimi K3 with substantially more post-training compute, centred on large-scale reinforcement learning (RL) across hundreds of downstream domains. The post calls it "3T-class" (exact size and architecture are not disclosed); context is 1M tokens.
Availability. Kimi.ai and mobile apps, Kimi Work, Kimi Code and the Kimi API (IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed). Full weights are promised by November 4, 2026; the post does not state a license.
Company-reported numbers (max reasoning effort unless noted):
- Terminal-Bench 4.0: 65.7% vs 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5.
- Kimi Code Bench 3.0 Long-horizon (internal): 61.2% vs 55.95% for K3.
- BenchCAD (Voxel IoU, 1,000-file subset) with tools: 92.7 vs 96.2 for Claude Opus 5.5 and 95.9 for GPT-6 Astra.
- Kimi Hallu Bench Hard: claim-level hallucination rate fell from 23.4% (K3) to 8.3% (K3.1 Pro with tools).
- Internal finance and business tests show large gains over K3 (e.g. Business Operation: 37 vs 8.9).
Caveats. Nearly all numbers are Moonshot's own, several benchmarks are internal, and there is no independent verification yet. On xdev-bench-hard the company reports 81.0% for the released model and 90.5% for a version without cyber risk-reduction measures, so the safety restrictions noticeably cut performance in that area.
Moonshot AI is a Chinese company that makes the Kimi chatbot. It has released a new "brain" for that bot: the K3.1 Pro model.
Why it matters:
- It can keep a huge amount of text in mind — roughly several thick books at once. That helps when working through a big software project or a long report.
- It writes code well. On one test where the model works like a programmer at a command line, it scored almost as high as the strongest current model from Anthropic.
- It is promised to become open. The strongest models are usually kept private and offered only through a website. Moonshot says it will publish the model's "weights" — the model itself — by November 4, so anyone could run it on their own servers.
Keep in mind: the company measured all of these results itself. Independent tests will come later, and the numbers may turn out more modest.