Moonshot releases Kimi K3.1 Pro through hosted services, with weights promised next

Moonshot AI is offering Kimi K3.1 Pro through its apps and developer services, but the downloadable release remains a promise. Unite.AI reports that the new model and a faster Highspeed variant are available across Kimi’s services. Moonshot says full weights will follow by November 4, 2026, alongside company-reported gains in coding, financial analysis and factual accuracy.

Hosted first, with access tied to membership

K3.1 Pro is the first model in Moonshot’s K3.1 series and a further-trained version of Kimi K3. Moonshot says it used substantially more post-training computation, improving systems for reinforcement learning, long-context processing and running tasks in isolated environments. The company describes it as a 3T-class model with a context window of up to one million tokens—the material it can process in a request.

Access spans Kimi.ai, mobile apps, the Kimi Work desktop app, Kimi Code and the API. Moonshot says Highspeed delivers identical capabilities at two to three times the speed. In Kimi Code, however, the faster option uses about twice the quota. Its documentation also ties model access and the largest context window to membership tiers:

• Under new membership plans, K3.1 Pro requires Plus or above; the one-million-token window requires Pro or above.
• Highspeed requires Pro or above. Both models accept images and video, but support thinking mode only; requests that disable it are rejected.
• The documentation recommends starting a fresh Kimi Code session. A mistyped model identifier in third-party tools falls back to K2.8 Preview without thinking.

Enterprise administrators can enable K3.1 Pro for their organizations, but access is off by default at launch. API customers also have a Zero Data Retention option.

Close coding scores, a narrower accuracy claim

Moonshot’s headline coding comparison puts K3.1 Pro at 65.7% on Terminal-Bench 4.0, which evaluates realistic terminal tasks under resource limits. That is just below Claude Opus 5.5 at 66.4% and above GPT-6 Astra at 58.2% in the company’s comparison. These are Moonshot-reported results, not an independently established ranking.

The company also reports 92.7% with tools on BenchCAD, a test of reconstructing 3D objects from multiple views by generating CAD code. Its comparison puts Claude at 96.2% and GPT-6 Astra at 95.9%. For financial-document analysis, Moonshot reports 64.4 on OfficeQA Pro v2 and describes that result as state of the art.

The factual-accuracy claim needs a different reading. Moonshot’s internal Kimi Hallu Bench Hard deliberately probes situations likely to produce errors. The company says K3.1 Pro sharply reduced the share of individual claims containing hallucinations, but explicitly warns that these figures do not represent error rates in everyday use.

A deliberate limit before downloadable weights

Moonshot says stronger post-training also improved vulnerability research, tracing that capability to cyber-related material in mid-training data and reinforcement-learning tasks. For the open-weight release, the company says it removed cyber-related data from both stages while keeping general-purpose coding tasks. That is a training intervention, rather than simply an access restriction on the hosted service.

On its internal xdev-bench-hard vulnerability test, Moonshot reports 90.5% for a checkpoint without those measures, versus 81.0% for the final model, with up to three attempts. The company argues that this meaningfully reduces cyber capability while leaving it behind proprietary frontier models publicly available for at least three months.

Its stated policy has two parts: avoid releasing weights that advance the absolute frontier of cyber capability, but progressively improve what open-weight models can do. Moonshot argues that broad access also helps teams audit their own software, including legacy code that rarely gets dedicated security reviews.