Moonshot AI has released Kimi K3.1 Pro, a 3-trillion-parameter open-weight model with a 1-million-token context window, available today on Kimi.ai, Kimi Code, and the Kimi API. Full model weights will be released by November 4.
The launch puts a new open-weight contender into a market dominated by GPT-6 Astra and Claude Opus 5.5. On Terminal-Bench 4.0, which tests coding agents on realistic terminal tasks, K3.1 Pro scores 65.7 percent, ahead of GPT-6 Astra at 58.2 percent and just behind Claude Opus 5.5 at 66.4 percent.
The timing matters because open-weight models have been closing the gap with proprietary systems all year. Stanford's 2026 AI Index found Anthropic's top model leads by just 2.7 percent as of March. Kimi K3.1 Pro extends that trend with a model users can run on their own infrastructure.
The 1-million-token context window is the other headline figure. That is enough to feed the model an entire codebase, a long financial filing, or a book-length document in a single prompt. For agentic workflows that need to hold state across many steps, that headroom reduces the need to summarize or chunk inputs.
What the benchmark numbers show
K3.1 Pro is a post-trained version of Kimi K3, built with substantially larger post-training compute. Moonshot optimized every layer of its stack, from the reinforcement learning framework to sandbox infrastructure, so that each increment of compute delivers stable returns. The company claims consistent gains across every domain it tested.
On BenchCAD, which tests 3D reconstruction from multi-view renderings through CAD code generation, K3.1 Pro scores 92.7 percent with tools. That approaches Claude Opus 5.5 at 96.2 percent and GPT-6 Astra at 95.9 percent. On Kimi Code Bench 3.0 Long-horizon, an internal benchmark for multi-step coding, it scores 61.2 percent, up from K3's 55.95 percent.
The model also covers the full range of coding work, from quick fixes to full-project builds. It works across Unity, Unreal Engine, and Godot, combining web search and 3D generation to create complete experiences. Moonshot says its front-end pages and interactive scenes show stronger composition, typography, and animation than the previous version.
In one internal test, K3.1 Pro built a human walking dynamics simulator on top of human-atlas, an open-source anatomical model with 296 skeletal components and 402 muscle components. Across 167 missions and 420 independent reviews, it analyzed results, diagnosed problems, and iteratively refined the controller. When reward shaping failed to improve arm motion, it ran open-loop probes of the simulator's dynamics, identified and corrected 36 errors in shoulder-muscle moment-arm calculations, and retrained the controller. The simulated human went from collapsing within seconds to walking continuously for 11.2 seconds.
Finance and hallucination numbers
In finance, K3.1 Pro achieves state-of-the-art results on OfficeQA Pro v2 at 64.4, which tests complex fiscal data analysis over real documents. It also posts gains across internal finance benchmarks: Financial Modeling improves by 7.2 points, Financial Research by 11.6 points, and Business Operation reaches roughly 4.2 times K3's performance.
Factual accuracy improved too. On Kimi Hallu Bench Hard, an internal stress test for factual accuracy, the claim-level hallucination rate fell from 23.4 percent for K3 to 8.3 percent for K3.1 Pro. The share of responses containing at least one factual error dropped to 42.4 percent of K3's rate. Moonshot notes these are stress-test figures, not overall daily error rates.
Cybersecurity controls and open-weight principles
Moonshot spent a larger share of effort on cybersecurity as the model grew more capable. The company identified and removed cyber-related data from both mid-training and reinforcement learning stages for the open-weight release. On xdev-bench-hard, a benchmark of vulnerability discovery and exploitation tasks, the checkpoint without these measures solves 90.5 percent of problems at pass@3, while the final K3.1 Pro reaches 81.0 percent.
K3.1 Pro's cyber capability remains behind frontier proprietary models that have been available for at least three months. Moonshot frames this as an intentional choice: the company says it aims not to extend the absolute frontier of cyber capability through open weights, and instead wants the capability to diffuse gradually. The company argues that running a capable model on infrastructure users control lets more teams audit their own code, including legacy systems that rarely get dedicated security reviews.
A high-speed variant and 3D tools
Moonshot is also releasing Kimi K3.1 Pro Highspeed, which delivers two to three times the speed of K3.1 Pro for more responsive interactions. Both versions are available today on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
3D Make is now live on kimi.ai and the Kimi app, letting users preview and edit 3D models online, download them for 3D printing, or export them as videos. Enterprise access is off by default at launch; administrators can enable it.
How Moonshot positions the release
Moonshot says K3.1 Pro is the first model in a new K3.1 series. The company is pushing the open-weight angle hard, with weights coming November 4 and API access today. That puts pressure on other open-weight providers to match the combination of a 1-million-token context window and frontier-level coding scores.
The Highspeed variant matters for product teams. Two to three times the throughput at the same quality level changes the economics of running agents in production. For latency-sensitive applications like coding assistants, that speed difference is the feature, not the benchmark score.
Moonshot also published detailed evaluation methodology, including 12 footnotes on benchmark conditions. That level of transparency is unusual and makes the numbers easier to check, though it also means the headline scores depend on specific harness choices like temperature settings and step limits.
What to watch next
The full weights release by November 4 will determine whether the open-source community can reproduce Moonshot's benchmark claims. It is not clear this will hold once independent labs run their own tests. The model's cyber capability tradeoff, with offensive scores held below what Moonshot's own pipeline could produce, will also face scrutiny from researchers who want to know whether the data removal actually limits real-world misuse.
For developers, the practical question is whether a 3-trillion-parameter model with a 1-million-token context window runs well enough on consumer hardware to matter. Moonshot has not published minimum hardware requirements. Until then, the API and Kimi Work desktop app, version 3.1.0 or later, are the practical entry points.
Read more about the Kimi K3.1 Pro tech blog and browse AI coverage on jualin.