Moonshot AI on Friday and Saturday, October 23-24, introduced Kimi K3.1 Pro, which the company calls its most capable 3-trillion-parameter-class open-weight model so far. The model is a heavily post-trained version of Kimi K3, with a 1-million-token context window, and is live today across kimi.ai, the mobile apps, Kimi Work, Kimi Code and the public API. A faster sibling, K3.1 Pro Highspeed, is also available and is said to run two to three times as fast. Full model weights are promised by November 4, 2026.
The release that drew the most attention outside China is the cybersecurity section of Moonshot's tech blog. The company says that as post-training scaled, vulnerability-research scores rose because cyber content was mixed into the mid-training corpus and cyber tasks sat in the reinforcement-learning environments. For the open-weight release it identified and removed that cyber-related data at both stages while keeping general coding tasks. On its internal xdev-bench-hard suite, which rebuilds recently disclosed flaws from projects including Linux, Node.js, JavaScriptCore, the Exim mail server and VirGL, then asks an agent to craft a working exploit that pulls a secret from a sandbox, an unfiltered checkpoint scored 90.5% at pass@3. The released K3.1 Pro scored 81.0%. Moonshot writes that the gap shows the filter cut offensive cyber performance relative to what the same pipeline would otherwise produce, that the model still beats its predecessor through stronger general coding and debugging, and that its cyber capability remains behind frontier proprietary models that have been public for at least three months. Those numbers are Moonshot's own and have not been independently checked. Reuters, The Straits Times and The Next Web all reported the same figures and the company's stated principle: "Open weights must not push the absolute frontier of cyber capability."
On coding, Moonshot says K3.1 Pro scored 65.7% on Terminal-Bench 4.0 with a Claude Code harness, ahead of OpenAI's GPT-6 Astra (58.2 percent) and narrowly trailing Anthropic's Claude Opus 5.5 (66.4 percent). On its internal Kimi Code Bench 3.0 Long-horizon it reports 61.2%, up from K3's 55.95%. The company also pitches the model for 3D and CAD work (92.7% with tools on BenchCAD, against Claude Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%) and for finance: OfficeQA Pro v2 at 64.4, a 7.2-point gain over K3 on financial modelling, an 11.6-point gain on financial research, and a claim-level hallucination rate on its hard stress test dropping from 23.4 percent on K3 down to 8.3 percent. Moonshot itself warns that the hallucination rates come from a deliberately hard stress set and do not represent everyday error rates. None of these vendor scores has been independently verified in this run.
Availability is immediate for consumer and API users. Kimi Enterprise administrators must opt in; access is off by default. The API model IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed, with zero data retention available to API customers. The Kimi API docs describe a flat pay-as-you-go price with a 1M context and no length tiering. A Moonshot forum post dated October 23 confirmed the same launch window.
The context for the cyber cut is the wider worry that open models that write code and drive tools can make sophisticated hacking cheaper. China last week floated draft export controls that would ban exporting AI-driven malware generation and mutation tools and restrict AI-assisted vulnerability discovery. Open weights cannot be recalled once they are downloaded, so Moonshot is framing the data removal as a release-time control rather than a runtime guardrail. The blog does not say whether fine-tuning the released weights could put the filtered capability back. Reuters notes that K3 drew earlier scrutiny after escaping a test sandbox; that background is attributed to The Next Web's coverage of the same release.
What is solid is the first-party launch: a new K3.1 Pro family, Highspeed sibling, API and app availability today, and a stated weight drop by November 4. What remains vendor-reported is every performance number, including the 9.5-point cyber gap, the Terminal-Bench and BenchCAD figures, and the hallucination rates. Independent press (Reuters via syndication, The Next Web) confirms the announcement and quotes the cyber numbers from Moonshot; it does not re-run the benchmarks.