On 23 October 2026, a Beijing startup did something the biggest AI labs in the world have mostly stopped doing: it announced a frontier-class model and promised to hand the whole thing away. Moonshot AI’s Kimi K3.1 Pro is a 2.8-trillion-parameter open-weight model that, on Moonshot’s own numbers, sits within a whisker of Anthropic’s Claude Opus 5.5 on agentic coding — and the company says the full model weights will be public by 4 November. If that release happens as promised, it will be the strongest open-weight model available, and it will change the arithmetic for any business that builds on AI.

That “if” matters. Nearly every headline number below comes from Moonshot itself, not from independent testing — independent testers get their real chance only once the weights ship. What follows is what is confirmed, what is claimed, and what it means for you either way.

What Moonshot actually launched

Kimi K3.1 Pro is the first model in Moonshot’s new K3.1 series, and the company describes it as its most capable 3T-class open-weight model to date. It is not a brand-new base model. It is a post-trained version of Kimi K3, which Moonshot released in July 2026, put through substantially more reinforcement learning across hundreds of downstream domains.

The architecture is a mixture-of-experts design: roughly 2.8 trillion parameters in total, about 104 billion active per token — only running a slice of itself per request is how a model this large can be served at all. It also carries a 1-million-token context window — enough for a sprawling codebase or a shelf of financial documents in a single pass — and handles text and image inputs.

Two versions went live at once: Kimi K3.1 Pro itself, available now in the Kimi app, Kimi Work, Kimi Code and the Kimi API; and Kimi K3.1 Pro Highspeed, which Moonshot says delivers the same capabilities at two to three times the speed for interactive work where latency matters most.

The benchmark story, honestly told

Start with the caveat: the benchmark scores below are Moonshot’s own. The company ran its model inside the Claude Code harness for most coding tests and took rival scores from published reports or official leaderboards — which is why independent verification after 4 November will matter so much.

With that caveat, the pattern is consistent enough to take seriously. On Terminal-Bench 4.0, which grades coding agents on realistic terminal tasks, Moonshot reports 65.7% for K3.1 Pro — ahead of OpenAI’s GPT-6 Astra at 58.2% and just 0.7 points behind Claude Opus 5.5 at 66.4%. On SWE-Marathon, Moonshot says K3.1 Pro leads at 51.7%, ahead of Opus 5.5’s 46.3%, and on its internal long-horizon test, Kimi Code Bench 3.0, the score improved from 55.95% for K3 to 61.2%.

Independent benchmarking adds a second data point. Artificial Analysis puts K3.1 Pro fourth on its Intelligence Index with a score of 55 — behind only Claude Opus 5.5, Claude Sonnet 5.5 and Muse Wonder Full, and ahead of GPT-6 Astra. On its GDPval-AA test of real-world work tasks, the model scored 1,846 Elo, up 312 Elo from Kimi K3 and third overall. Notably, Artificial Analysis’s own Terminal-Bench 4.0 run recorded 67.2%, slightly higher than Moonshot’s 65.7% — a reminder that different harnesses produce different numbers.

The picture is not uniformly rosy, and Moonshot’s own charts admit it. On Terminal-Bench Science, K3.1 Pro scores just 38.1%, far behind GPT-6 Astra’s 68.1% and Opus 5.5’s 63.3%. On BenchCAD, which asks models to rebuild 3D objects as CAD code, it trails both frontier rivals even though it leads Claude Fable 5.1. For heavy scientific computing, the closed frontier models still hold a clear edge.

Two other claimed improvements deserve attention. On Moonshot’s internal hallucination stress test, the claim-level error rate fell from 23.4% for K3 to 8.3% — a large drop, though again a company-run test. And on OfficeQA Pro v2, a test of analysis across real financial documents, Moonshot claims the top score of any model at 64.4, fitting its pitch that K3.1 Pro can handle end-to-end financial workflows.

The unusual part: a model deliberately made worse at hacking

The most interesting decision in this launch is not a benchmark. It is a subtraction. Moonshot says it removed cyber-related data from the training corpus and reinforcement-learning environments used to develop K3.1 Pro, while keeping its general coding skills intact. The company stated it does not aim to release open weights that would extend the absolute frontier of cyber capability.

The numbers show what that trade-off cost. On Moonshot’s internal exploit benchmark, the filtered K3.1 Pro scored 81.0%, against 90.5% for an unfiltered version of the same model trained with the cyber data left in. So the filtering shaved roughly nine and a half percentage points off its offensive performance. But the comparison with the previous generation is starker: the earlier Kimi K3 scored just 33.3% on the same test. Read plainly, the filtered model is still far more capable at exploitation than its predecessor. Moonshot slowed the climb; it did not freeze it.

Why does this matter? Because open weights change the control equation: with a hosted, closed model, the developer can monitor use, update safeguards, and revoke access. Once weight files are widely distributed, the original developer has far less ability to restrict access or update safeguards. Moonshot’s filtering is an attempt to manage that reality before the release, not after it — at a moment of growing worry that agentic AI could make sophisticated hacking easier and cheaper.

It also lands in a shifting regulatory frame. China has proposed draft export controls covering AI-driven malware generation and AI-assisted vulnerability discovery tools, with a temporary pre-export registration requirement already in force. That context does not change what K3.1 Pro can do, but it shapes how freely open-weight releases of this power will circulate.

What “open weights” really means for your business

Open weights are not the same as open source, and the distinction is worth getting right before you plan around it. When the weights ship, developers will be able to download the files, run the model on their own infrastructure, fine-tune it, and modify it. What they will not necessarily get is the training data, training code, or reinforcement-learning environments needed to reproduce the model from scratch.

Even with that limit, the practical consequences are large:

• Cost control. Running a model yourself, or through a smaller host, breaks the link between your usage and a single vendor’s price list. Moonshot has already cut API pricing by about a third versus K3 — $2 per million input tokens and $10 per million output tokens — and competitive hosting should push effective prices down further once weights are public.
• Data residency. A model you run on your own servers keeps prompts and documents inside your boundary — often the difference between “we can use AI for this” and “we cannot” for law firms, clinics, banks and government suppliers.
• Customisation. Fine-tuning on your own support tickets, codebase or product documentation produces a specialist a general hosted model cannot match — and you keep the result.
• No single point of failure. A vendor price rise, policy change or outage does not strand your product when you hold the weights — the core strategic argument for open weights, and why this release matters beyond its benchmark scores.

The trade-offs are equally real: a 2.8-trillion-parameter model is not something you run on a laptop, so most businesses will still use a hosted version. For most teams, open weights mean independence from a single vendor, not from infrastructure altogether — and a model you host is a model you must secure, update and monitor yourself.

Where to try it, and what to test first

K3.1 Pro is live now in the Kimi app and on the web, in Kimi Work, in Kimi Code, and through the Kimi API, with a limited free tier. If you want to evaluate it, start where the evidence is strongest:

• Agentic coding on a real repository. Give it a multi-file task with tests, not a single-function prompt — Terminal-Bench-style tasks are where the model claims its edge. Measure pass rates against whatever you use today.
• Long-document work. Feed it a large financial report, a contract set or a full codebase and ask for analysis connecting facts across distant sections — the 1-million-token window is the feature most likely to change a workflow.
• 3D and design generation. BenchCAD performance suggests the model can turn descriptions into CAD code and animate models — worth a direct test if you prototype products.
• Hallucination checks. Run your own factual-accuracy test on your own documents before trusting outputs in customer-facing work — Moonshot’s 8.3% internal figure is encouraging, not a guarantee.

What not to do yet: build a production system that depends on the open weights before they ship. The 4 November date is a promise, not a delivery — wait for the files, the licence terms and the first independent reproductions before making architectural commitments.

What this launch does not settle

Three open questions should temper any excitement. First, independent verification is still pending: Moonshot’s numbers are plausible and partly corroborated by Artificial Analysis, but the full picture arrives only when independent labs run the released weights themselves.

Second, the licence and usage terms for the weights are not yet public. “Open weights” covers a wide range, from permissive licences to terms with commercial restrictions. Read the actual licence before assuming you can ship a product on top of the model.

Third, the safety debate is unresolved, not settled, by Moonshot’s filtering. A model scoring 81% on an internal exploit benchmark is still a powerful dual-use tool, and the open-weights genie does not go back in the bottle. Expect regulation to keep moving, in Beijing, Washington and Brussels alike.

Practical takeaways

• If you pay for frontier coding models, test K3.1 Pro head-to-head on your own tasks before your next renewal — especially for agentic, multi-step coding work.
• If data residency blocks your AI adoption, mark 4 November: the weights release, not the hosted launch, is the event that matters for you.
• If you build products on AI, price in a world where a near-frontier model is downloadable — vendor lock-in is getting weaker as a strategy and stronger as a risk.
• Treat all launch benchmarks as claims until independent labs reproduce them on released weights. Budget a verification sprint, not just an adoption sprint.

Frequently asked questions

Is Kimi K3.1 Pro free to use?

Yes, with limits: a free tier with limited usage through the Kimi app and website, with full model weights promised by 4 November 2026. API access is paid, starting at $2 per million input tokens.

Is Kimi K3.1 Pro better than Claude Opus 5.5?

On some tests, nearly. Moonshot reports 65.7% on Terminal-Bench 4.0 against Opus 5.5’s 66.4%, and a lead on SWE-Marathon, while Opus 5.5 still leads on scientific computing and 3D CAD. The honest answer is “comparable on coding, behind on science” — and all figures await independent confirmation on the released weights.

What does “open-weight” mean?

It means the trained model files will be publicly downloadable, so developers can run, fine-tune and modify the model on their own systems — not necessarily including the training data or training code.

When will the model weights be released?

Moonshot says by 4 November 2026 — until the files and licence terms are actually published, treat that date as a promise rather than a fact.

Should my business switch to Kimi K3.1 Pro?

Not on launch headlines alone. Test it on your own workloads, check the licence when the weights ship, and weigh hosting costs and data-residency needs. For teams blocked by data residency or vendor lock-in, it is the year’s most important open release to evaluate.

Conclusion

Kimi K3.1 Pro is two stories at once. The first is capability: a post-trained open-weight model that, on the evidence so far, challenges closed frontier models on agentic coding. The second is governance: a lab deliberately dulling its own model’s cyber-offensive edge before giving it away, in a world where giving powerful models away is itself becoming a geopolitical act.

Together, they make 4 November a date worth marking. If the weights ship as promised with workable licence terms, the gap between the best open and best closed models narrows from a chasm to a few points — and the strategic question for every business using AI shifts from “which vendor do we rent intelligence from?” to “what do we build now that near-frontier intelligence is downloadable?” That is a question worth answering carefully, with your own tests, not someone else’s benchmarks. If you are exploring AI tools for your work, our guides to free online tools, AI chat platforms and free AI tools for students are a practical place to start — and for the policy backdrop, see our coverage of the US order removing DeepSeek from federal systems.