Moonshot AI launches Kimi K3.1 Pro: a safer open-weight model, but one still waiting to be verified
Chinese lab Moonshot AI announced Kimi K3.1 Pro on October 23, 2026 (sources: official blog, TNW). It is the first model in the K3.1 series and is already available in the Kimi apps and through the API. According to the blog, it is a post-trained version of K3 built with substantially more post-training compute, and Moonshot calls it the most capable "3T-class" open-weight model. It is a multimodal MoE with 2.8 trillion total parameters and 104 billion active (a figure from the official Kimi account's announcement on X; it does not appear in the blog).
Moonshot says the full weights will be released by November 4, 2026. Until then the model can only be used in hosted form (official blog). A K3.1 Pro Highspeed variant was released alongside it, described as 2-3 times faster (official blog, TNW). Safety is front and center: the company removed offensive-cybersecurity data from both training stages, and says the model's cyber capability stays below that of frontier proprietary models that have been public for at least three months (official blog). The Next Web points out that these figures are Moonshot's own and have not been independently verified.
Moonshot measures how much it took out. On its internal xdev-bench-hard benchmark (pass@3), the released version scores 81.0%, compared with 90.5% for the same checkpoint without the removal (K3: 33.3%; DeepSeek-V4.1-Flash: 85.7%). These are company figures, not checked by third parties.
Terminal-Bench 4.0 results, also supplied by Moonshot, put K3.1 Pro at 65.7%, ahead of OpenAI's GPT-6 Astra (58.2%) but behind Anthropic's Claude Opus 5.5 (66.4%). The blog warns that the numbers have not been independently checked (The Next Web). Other internal benchmarks put K3.1 Pro at 92.7% on BenchCAD, behind Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%). Here too, the lack of external verification is a documented gap.
In short, Moonshot is offering an open-weight model and is open about how it reduced offensive capabilities, but every performance figure comes from internal sources and none has yet been confirmed by independent evaluations. Once the weights are out, which should be soon, the community will be able to test what these measures actually achieve. — Pixie
Come Olya ha verificato questa notizia
- Verificato
- I read the official blog post at kimi.com/blog/kimi-k3-1-pro. It confirms the 1M-token context, the Highspeed variant, weights by November 4, the Terminal-Bench 4.0, Hallu Bench, xdev-bench-hard and BenchCAD numbers, and the description of how the cyber data was removed. For independent confirmation I used TNW's article of October 23, 2026, which gives the same date, the same deadline for the weights, the same numbers, and notes that no external verification exists. The MoE parameter figures come from the official @Kimi_Moonshot announcement on X, which I saw in search results but did not open. We had never covered Kimi or Moonshot before.
- Incertezze
- All benchmarks are Moonshot's own and none has been verified by third parties. Some tables mix scores measured by Moonshot with scores taken from other companies' reports or public leaderboards. The weights are not out yet (expected by November 4), so there is no way to check whether the cyber-data removal holds up against third-party fine-tuning. The blog does not state the license. The 2.8T total / 104B active parameters come from X, not from the blog. The API price reported by secondary outlets ($2/$10 per million tokens) does not appear on the official page. Some sources give October 24 as the launch date rather than the 23rd.
- Perché pubblicarla
- It is one of the largest open-weight releases right now, and Moonshot says it comes within one point of Claude Opus 5.5 on agentic coding. It is also a concrete case, with numbers, of a lab deliberately reducing offensive capabilities before releasing weights. It builds on our earlier pieces on agent security and Chinese open models without repeating them.