Moonshot's Kimi K3.1 Pro adds 1M-token context to open weights, plus a Highspeed variant

Moonshot AI has released Kimi K3.1 Pro, the first model in its K3.1 series and, by the company's own account, its strongest open-weight model in the 3-trillion-parameter class so far. The model ships with a 1-million-token context window and is joined by Kimi K3.1 Pro Highspeed, a variant tuned to run 2-3 times faster for interactive use. Both are available now through Vercel's AI Gateway, according to Vercel, with the full model weights promised by November 4.

A larger post-training run on top of Kimi K3

According to Moonshot's technical blog, K3.1 Pro is a post-trained version of Kimi K3 that consumed substantially more post-training compute. To run joint reinforcement learning across hundreds of downstream domains, the team says it reworked the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling and environment quality checks, so that each extra unit of RL compute produces stable, predictable gains.

Moonshot reports that the result matches or comes close to the strongest proprietary models on coding, agentic and knowledge benchmarks. On Terminal-Bench 4.0, which tests coding agents on realistic, resource-constrained terminal tasks with strict tests, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra at 58.2% and narrowly behind Claude Opus 5.5 at 66.4%. On Kimi's internal Code Bench 3.0 long-horizon evaluation of multi-step execution and tool use, it reaches 61.2%, up from K3's 55.95%.

Coding, 3D work and financial analysis

Moonshot says K3.1 Pro spans the full range of coding work, from small fixes in existing codebases to whole-project builds. It works across the Unity, Unreal Engine and Godot game engines, pairs them with web search and 3D generation, and produces front-end pages and interactive Three.js scenes with better composition, typography and animation than earlier Kimi models.

The company cites one internal long-horizon test in which the model built a human walking dynamics simulator on top of human-atlas, an open anatomical model with 296 skeletal and 402 muscle components. Across 167 missions and 420 independent reviews, it diagnosed problems and refined the controller until the simulated figure walked continuously for 11.2 seconds; when reward shaping stalled, it probed the simulator's dynamics directly, found and corrected 36 errors in shoulder-muscle moment-arm calculations, and retrained.

On BenchCAD, which tests 3D reconstruction from multi-view renderings via CAD code generation, K3.1 Pro scores 92.7% with tools, approaching Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%). Moonshot highlights its parametric modeling, noting it can write and run scripts for tools like Blender and CAD software without MCP connectors. A new 3D Make feature is live on kimi.ai and the Kimi app for previewing, editing, 3D-print-ready downloads, and export as videos or Live Photos.

In finance, Moonshot reports state-of-the-art results on OfficeQA Pro v2 (64.4), which tests analysis of real financial documents, plus internal gains over K3: 7.2 points on Financial Modeling, 11.6 points on Financial Research, and roughly 4.2x on Business Operation. On Kimi Hallu Bench Hard, an internal hallucination stress test, claim-level hallucination fell from 23.4% to 8.3%, and responses containing at least one factual error dropped to 42.4% of K3's rate. Moonshot cautions the test uses deliberately hard prompts and does not represent everyday error rates.

Cyber capability deliberately reduced for the open release

Moonshot says it noticed vulnerability-research performance climbing as post-training scaled, traced this to cyber-related content in the mid-training corpus and cyber tasks in the RL environment mix, and removed that data from both stages for the open-weight release while keeping general-purpose coding tasks. It evaluated the outcome on xdev-bench-hard, its in-house benchmark of expert-curated, source-code-based vulnerability discovery and exploitation tasks.

Available now through Vercel AI Gateway

According to Vercel, both models keep the 1M-token context window and image input from K3. They can be called through the AI SDK, the Chat Completions, Responses and Anthropic Messages APIs, or a coding agent connected to AI Gateway, using the model names moonshotai/kimi-k3.1-pro and moonshotai/kimi-k3.1-pro-highspeed. For Claude Code, Codex, Cursor and similar tools, Vercel recommends installing the latest CLI and running npx vercel ai-gateway setup. More providers will follow once the weights are out, and Vercel notes that AI Gateway passes through provider pricing with no markup or platform fee, including on Bring Your Own Key, alongside Zero Data Retention support, key budgets, routing rules, retries and failover.

Why it matters

A 3T-class open-weight model with a 1-million-token context that trades benchmark blows with proprietary flagships narrows the gap between open and closed AI, and the weight release on November 4 will let developers run it themselves. The Highspeed variant acknowledges that raw capability is not enough for interactive agent workloads, where latency matters. Just as notable is the security disclosure: Moonshot is documenting a deliberate decision to strip cyber-offensive training data before open-sourcing, an approach other labs releasing weights may come under pressure to copy. Day-one availability through a gateway with pass-through pricing lowers the friction for teams that want to try the model before committing to self-hosting.