Moonshot AI launches Kimi K3.1 Pro, a 3-trillion-parameter open-weight model

Chinese AI lab Moonshot AI has released Kimi K3.1 Pro, the first model in a new K3.1 series and, according to the company, its most capable open-weight model to date at roughly 3 trillion parameters. A faster variant, Kimi K3.1 Pro Highspeed, ships alongside it, running 2 to 3 times faster than the standard model. Full open weights for both are scheduled for release by November 4.

The model is post-trained from the earlier Kimi K3, released in July, using what Moonshot describes as substantially more reinforcement-learning compute spread across hundreds of domains. It supports a 1-million-token context window and is available now through kimi.ai, the Kimi mobile apps, the Kimi Work desktop app (version 3.1.0 and above, on Windows and Apple Silicon Macs), and Kimi Code via the /model command. Developers can access it through the Kimi API under the model IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed, with zero data retention offered as an option for API customers. Enterprise and business customers can enable the model through an admin setting; it is off by default at launch.

On Moonshot's published benchmarks, K3.1 Pro scored 65.7% on Terminal-Bench 4.0, trailing Anthropic's Claude Opus 5.5 at 66.4% but ahead of GPT-6 Astra at 58.2%. On Kimi Code Bench 3.0's long-horizon coding tasks, it scored 61.2%, up from 55.95% for the prior K3 model. On BenchCAD, a 3D-reconstruction benchmark, it reached 92.7% with tool use (versus 96.2% for Opus 5.5 and 95.9% for GPT-6 Astra) and 72.8% without tools, roughly matching Opus 5.5's 73.0%. The company also reported state-of-the-art results on its internal OfficeQA Pro v2 benchmark (64.4) and gains on finance-specific tests: financial modeling rose from 50.4 to 57.6, financial research from 58.8 to 70.4, and business-operation tasks jumped from 8.9 to 37 compared with K3.

Moonshot says it also made progress on reliability: on its Hallu Bench Hard test, which measures claim-level hallucination rates, K3.1 Pro's error rate with tool access fell from 23.4% for K3 to 8.3%. The company separately disclosed that it removed cyber-related training data from the mid-training and reinforcement-learning stages before releasing the open weights, a step it says reduces offensive cyber capability — K3.1 Pro scored 81.0% on the xdev-bench-hard cyber benchmark, down from 90.5% for an internal checkpoint without that mitigation. Moonshot states the released model's cyber capability remains behind frontier proprietary systems.

Alongside the model, Moonshot launched a new “3D Make” feature on kimi.ai and in the Kimi app, letting users preview, edit, and export 3D models generated by the AI. Pricing for API access was not detailed in the announcement; Moonshot pointed users to its separate pricing page.

The release lands as open-weight Chinese models continue to close the gap with closed, US-built frontier systems on coding and agentic benchmarks, while competing primarily on cost and availability for developers who want to self-host or fine-tune rather than call a proprietary API.