--- title: "Kimi K3.1 Pro nears Claude Opus 5.5 on coding, weights by 4 November" url: https://blog.vomini.com/blog/kimi-k3-1-pro-open-weights date: 2026-10-26T13:30:00+00:00 updated: 2026-10-28T01:47:19.531961+00:00 author: "Roberto R." section: "News" language: en site: "Vomini" license: free to quote with attribution --- # Kimi K3.1 Pro nears Claude Opus 5.5 on coding, weights by 4 November > Moonshot released Kimi K3.1 Pro, which it calls its most capable open-weight model, with full weights promised by 4 November. On Moonshot's numbers it scores 65.7% on Terminal-Bench 4.0, just under Claude Opus 5.5… Moonshot AI, the company behind the Kimi models, released Kimi K3.1 Pro, which its blog lists on 24 October. Moonshot calls it the first model in its new K3.1 series and its most capable open-weight model to date, with a 1 million token context window. A faster version, K3.1 Pro Highspeed, runs two to three times quicker. Both are available on Kimi.com, in the Kimi Work and Kimi Code apps and through the Kimi API. Moonshot says the full model weights will be released by 4 November, so anyone will be able to download and run the model on their own servers. ## What Moonshot reports On Terminal-Bench 4.0, a test of coding agents working in a terminal, Moonshot reports 65.7% for K3.1 Pro, narrowly behind Claude Opus 5.5 at 66.4% and ahead of GPT-6 Astra, which it lists at 58.2%. OpenAI's own announcement gave Astra 57.9% on the same test, and Moonshot's footnotes say it took the highest figure reported anywhere for each model. On OfficeQA Pro v2, which tests analysis of real financial documents, Moonshot says K3.1 Pro leads with 64.4. Moonshot also says it worked on factual accuracy. On an internal stress test built to provoke mistakes, the share of false claims fell from 23.4% for Kimi K3 to 8.3%. Moonshot notes that this test is deliberately hard and does not reflect error rates in everyday use. ## A deliberate limit on cyber skills The most unusual section is about security. Moonshot says it found that cyber-related content in its general training data and training tasks kept raising the model's ability to find and exploit software flaws. For the open-weight release, it removed that data. On its internal exploit test, a version trained without the removal solved 90.5% of problems, and the final model solves 81.0%. Moonshot says the result stays behind the frontier closed models, and that it does not want its open weights to push the frontier of cyber capability. That is a different choice from the one Mistral made this month with [Mistral Large 4](/blog/mistral-large-4-preview-open-weights), which it pitched on its security strength. Moonshot is also the company Anthropic accused earlier this month of forwarding requests to Claude, an accusation we covered when Anthropic described its [defenses against model distillation](/blog/anthropic-distillation-defenses). Moonshot's post does not address it. ## Price and privacy The blog does not list API prices. It says zero data retention is available for API customers, and that for Kimi Enterprise organizations the model is off by default until an administrator turns it on. ## What to do this week If you are weighing open models for coding work, run the same set of real tasks through K3.1 Pro on the API now, so you can decide before 4 November whether to self-host the weights. Check where the API is hosted and under which law before you send it client code. ## What we think An open-weight model within a point of the best closed model on a hard coding test, on its maker's own numbers, is a big shift from a year ago. Independent results will tell whether it holds. The cyber section is worth reading in full: it is a rare case of a lab showing how much capability it chose to leave out. ## Sources - [Moonshot AI (Kimi)](https://www.kimi.com/blog/kimi-k3-1-pro), 24 Oct 2026 - [OpenAI, GPT-6 Astra announcement (for the Astra score)](https://openai.com/index/gpt-6-astra/), 3 Sep 2026 --- Originally published at https://blog.vomini.com/blog/kimi-k3-1-pro-open-weights