On October 23, 2026, Moonshot AI published Kimi K3.1 Pro, which it calls "our most capable 3T-class open-weight model to date, with a 1-million-token context window." It is "a post-trained version of Kimi K3, with substantially larger post-training compute", and a faster variant, K3.1 Pro Highspeed, runs at two to three times its speed, according to Moonshot. The API names are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. The release promises: "Full model weights will be released by Nov 4, 2026." The post does not name a licence.
The headline numbers
All figures below are Moonshot's own, and several benchmarks are internal.
- Terminal-Bench 4.0: 65.7%, ahead of GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%. Competitor scores are the maximum of their own blog posts or the official leaderboard, and Moonshot notes that "proprietary model testing may involve some fallback behavior."
- Hallucination stress test: on its Kimi Hallu Bench Hard, the claim-level hallucination rate fell from 23.4% for K3 to 8.3%. Moonshot adds that "The reported rates do not represent overall error rates in daily use."
- Office and finance work: 64.4 on OfficeQA Pro v2, which Moonshot describes as state-of-the-art for fiscal data analysis over real financial documents.
- CAD: 92.7% on BenchCAD with tools, below Opus 5.5 (96.2%) and GPT-6 Astra (95.9%).
The cyber disclosure
The most informative section is about what Moonshot chose to make worse. It says it traced a rise in offensive capability to the fact that "cyber-related content was embedded within the general mid-training corpus, and cyber-related tasks were present in the RL environment mix -- a natural byproduct of training on broad data." It responded by "Identifying and removing cyber-related data" at mid-training and filtering cyber RL environments. On its internal xdev-bench-hard, a checkpoint trained "no cyber-related data removal at mid-training and no cyber-related RL environment filtering" solves 90.5% of problems at pass@3, while "the final K3.1 Pro reaches 81.0%." In Moonshot's words, "data removal meaningfully reduces cyber capability relative to what the same training pipeline would otherwise produce." A nine-point drop on an internal benchmark is a modest effect, but it is a measured one. Moonshot's stated position is that "K3.1 Pro's cyber capability remains behind that of frontier proprietary models that have been publicly available for at least three months."
That is a specific, testable policy for an open-weight release: stay behind a moving proprietary frontier on offensive cyber, because weights cannot be recalled. It arrives in the same week that Anthropic described what an unsafeguarded frontier model did inside a cyber evaluation.
Deployment details
The model is available on Kimi.ai, the apps, Kimi Work, Kimi Code (via /model) and the API. "Zero Data Retention is available for API customers." Enterprise access is "off by default at launch" and must be enabled by an administrator. There are new 3D modelling features with Unity, Unreal Engine and Godot export.
What remains uncertain. Architecture details, active parameters, licence terms and pricing are not in the post. The lag-the-frontier cyber claim is self-assessed and will be checkable only once weights are public, and no independent lab has yet reproduced the Terminal-Bench score.
Analysis: If the weights ship on time, teams get a near-frontier agentic coding model they can host. Before switching, run your own terminal tasks, since agent benchmarks are sensitive to harness and timeouts, and price the hardware: a 3T-parameter model with 1M context is a multi-node deployment even if few parameters are active per token. The cyber filtering is the policy story. Data removal is one of the few safety measures that survives fine-tuning better than refusal training, and if other open-weight labs adopt and document it, external red teams finally get something concrete to verify.
Sources and evidence
Moonshot's Kimi K3.1 Pro (October 23) is a post-trained 3T-class model with a 1M-token context that reports 65.7% on Terminal-Bench 4.0 (Opus 5.5: 66.4%, GPT-6 Astra: 58.2%) and an 8.3% hallucination stress-test rate, with open weights promised by November 4; Moonshot says it removed cyber-related training data and RL environments so the model stays behind proprietary models released at least three months earlier, a self-assessed claim that can only be tested once weights ship.