Moonshot Apologises for Kimi K3.1 Pro Rate Limits and Resets Subscriber Quotas, With Weights Promised by November 4

Moonshot AI has apologised to Kimi subscribers for rate limits on its new flagship and is handing back a quarter of their monthly quota. In a post on X late on October 25 (UTC), early on October 26 in Asia, the company wrote: “We’re sorry for the rate limits and unnecessary quota consumption on the K3.1 Pro subscription plan. These were our issues, and we should have fixed them sooner.” Every subscriber on every plan gets a one-time reset that restores 25% of their monthly quota, which Moonshot equates to one week of use. It also says it changed the plan’s parameters to cut unnecessary subagent calls and shorten output.

The model behind the complaints, Kimi K3.1 Pro, was announced on October 23, 2026, as a post-trained version of Kimi K3. Moonshot calls it its most capable “3T-class” open-weight model. According to the company’s launch post on X, it has 2.8 trillion total parameters with 104 billion active per token in a mixture-of-experts design, a 1-million-token context window and multimodal input. It went live the same day on Kimi.ai, the Kimi apps, Kimi Work, Kimi Code and the Kimi API, alongside a Highspeed variant that Moonshot says runs two to three times faster. The weights are still a promise. “The full model weights will be released by Nov 4,” the technical blog says, but no Hugging Face repository exists yet, and no licence has been published.

Same base, much more reinforcement learning

Moonshot describes K3.1 Pro as K3 with “substantially larger post-training compute” rather than a new pre-training run. The blog credits most of the gains to reinforcement learning across hundreds of downstream domains at once. To get there, Moonshot says it reworked its RL framework, long-context inference, sandbox infrastructure and rollout scheduling so that extra RL compute produced predictable returns.

Every benchmark figure below is Moonshot’s own, and none has been independently reproduced. On Terminal-Bench 4.0, run through the Claude Code harness, the company reports 65.7% for K3.1 Pro, against 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. On its internal Kimi Code Bench 3.0 Long-horizon test the score rises to 61.2%, from 55.95% for K3. Moonshot also reports 92.7% on BenchCAD with tools, which tests rebuilding 3D objects as CAD code from multi-view renders. That trails the 96.2% Moonshot cites for Claude Opus 5.5 and 95.9% for GPT-6 Astra. On OfficeQA Pro v2, a financial-document benchmark, Moonshot claims a state-of-the-art 64.4. Unless a footnote says otherwise, the company ran its own model at maximum reasoning effort with temperature 1.0.

Game engines, Blender and spreadsheets

Much of the launch material is about work outside a code editor. Moonshot says K3.1 Pro builds full projects in Unity, Unreal Engine and Godot, and writes and runs scripts for Blender and CAD software to create and edit 3D assets without MCP connectors. A new “3D Make” feature on kimi.ai lets users preview models, download them for 3D printing or export them as video.

Finance gets its own section. Moonshot reports gains over K3 of 7.2 points on an internal financial-modelling benchmark and 11.6 points on financial research. On a simulated business-management test it reports about 4.2 times K3’s score. On its internal Kimi Hallu Bench Hard stress test, the blog says the claim-level hallucination rate fell from 23.4% for K3 to 8.3%. Moonshot cautions that the prompts are built to provoke errors and the rates don’t reflect everyday use.

Cyber data stripped before release

The blog’s longest section explains why the open-weight model is deliberately weaker at offensive security than it could be. Moonshot says vulnerability-research performance climbed steadily as post-training scaled. It traced that climb to cyber-related material in the general mid-training corpus and to cyber tasks in the RL mix, and removed both for the open-weight release.

Moonshot measured the effect on its own xdev-bench-hard, an in-house set of exploit-development tasks based on recently disclosed bugs in software such as the Linux kernel, Node.js and Exim. By Moonshot’s count, a checkpoint trained without the filtering solved 90.5% at pass@3 and the released K3.1 Pro solved 81.0%. Moonshot says the model still beats K3 on these tasks but remains behind frontier proprietary models that have been public for at least three months. Its stated policy is not to release open weights that “extend the absolute frontier of cyber capability”.

For developers, the API model IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed, and Moonshot offers zero data retention to API customers. Kimi Enterprise administrators have to switch the model on, since access is off by default. Until weights and a licence actually appear, K3.1 Pro is an open-weight model only by Moonshot’s promise; whether it becomes one in practice depends on the November 4 deadline being met and on the licence that comes with it.