The Pulse
Moonshot Launches Kimi K3.1 Pro With 1M-Token Context
Moonshot AI has released Kimi K3.1 Pro, a 3T-class open-weight model with a 1-million-token context window, alongside a faster Highspeed version. The company says the model improves on coding, 3D design and finance benchmarks, with full wei
AI.info Team · Written with AI · fact-checked ·
65.7% on Terminal-Bench 4.0
Moonshot AI says its new Kimi K3.1 Pro scored 65.7% on Terminal-Bench 4.0, a test of coding agents working through realistic terminal tasks. That puts it above GPT-6 Astra’s 58.2% and just below Claude Opus 5.5’s 66.4%, according to the company’s launch post. The comparison is Moonshot’s reported evaluation, not an independent assessment of the models.
Moonshot describes K3.1 Pro as the first model in a new series and its most capable 3T-class open-weight model to date. It has a 1-million-token context window and is a post-trained version of Kimi K3, with substantially more compute devoted to post-training. “It delivers substantial improvements across coding, 3D creation, financial research, and other agentic capabilities,” the company writes.
A faster version and a Nov. 4 weights date
K3.1 Pro is available starting Oct. 23 through Kimi.ai, the Kimi mobile apps, Kimi Work, Kimi Code and the Kimi API. Moonshot is also releasing Kimi K3.1 Pro Highspeed, which it says runs two to three times faster than the standard version. The company says the full model weights will be released by Nov. 4, 2026; until then, the launch offers access through its products and API rather than downloadable weights.
The release puts the model toward multi-step work rather than single-turn chat. On Moonshot’s internal Kimi Code Bench 3.0 Long-horizon evaluation, which measures coding agents’ tool use and end-to-end execution, K3.1 Pro scored 61.2%, compared with 55.95% for K3. The company also reports 92.7% on BenchCAD with tools, a test of generating CAD code from multiple views, against 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra.
Finance scores come with a stated caveat
Moonshot reports a score of 64.4 on OfficeQA Pro v2, which tests analysis of financial documents using retrieval, cross-document checks and numerical reasoning. Its internal finance evaluations show improvements over K3 of 7.2 points in financial modeling and 11.6 points in financial research; on its simulated business-operations test, K3.1 Pro scored about 4.2 times as well as its predecessor. The company says its figures generally use maximum reasoning effort, and these finance results are company-reported rather than independent comparisons.
Moonshot also flags a limit on how to interpret its hallucination testing. On its internal Kimi Hallu Bench Hard, the claim-level hallucination rate fell from 23.4% for K3 to 8.3% for K3.1 Pro, but the company says the benchmark deliberately stresses situations where models are prone to errors and does not represent everyday error rates. It also reports that 42.4% as many responses contained at least one factual error as in the predecessor’s results.
Moonshot says it removed cyber-related training data
The release includes a specific safety trade-off: Moonshot says it removed cyber-related data from both mid-training and reinforcement-learning stages while retaining general-purpose coding tasks. On its in-house xdev-bench-hard test, the final model solved 81.0% of problems at pass@3, compared with 90.5% for a checkpoint trained without those risk-reduction measures. The benchmark asks an agent to find and exploit vulnerabilities in supplied source code inside a separate sandbox.
That result is a comparison within Moonshot’s own test setup, not proof that the model is safe in all settings. The company says its aim is to keep K3.1 Pro’s offensive cyber capability below what its training pipeline could produce, while releasing weights for legitimate defensive work. The weights are scheduled for Nov. 4; the company’s post does not give a price for API use or specify what license will govern the downloadable release.