Moonshot AI launched Kimi K3.1 Pro on October 23, in the company’s words its “most capable 3T-class open-weight model to date.” The model carries a 1-million-token context window and arrives with a faster sibling, Kimi K3.1 Pro Highspeed, which Moonshot says runs two to three times faster. On Moonshot’s own Terminal-Bench 4.0 numbers, K3.1 Pro scores 65.7 percent, ahead of OpenAI’s GPT-6 Astra at 58.2 percent and within a point of Anthropic’s Claude Opus 5.5 at 66.4 percent. The release is already live across the Kimi product family, and Moonshot has committed to publishing the full model weights by November 4.
What K3.1 Pro Changes
K3.1 Pro is not a new base model. Moonshot describes it as a post-trained version of Kimi K3, the same 2.8-trillion-parameter mixture-of-experts architecture with 104 billion active parameters per token and text-and-image input, but pushed through substantially larger post-training compute: large-scale joint reinforcement learning across hundreds of downstream domains, with the RL framework, long-context inference and training efficiency rebuilt to make each increment of compute pay.
The headline result is an agent built for long-horizon work. On Kimi Code Bench 3.0 Long-horizon, Moonshot’s internal test of multi-step execution and tool use, K3.1 Pro scores 61.2 percent, up from 55.95 percent for K3. Demo work leans into full projects: games built end to end in Unity, Unreal Engine and Godot, and one test where the model iterated a walking-dynamics simulator across 167 missions and 420 reviews until the body walked for 11.2 seconds — fixing 36 shoulder-muscle calculation errors that reward shaping alone had missed.
Coding, 3D And Finance Benchmarks
On BenchCAD, which rebuilds 3D objects as CAD code, K3.1 Pro scores 92.7 percent with tools, against 96.2 percent for Claude Opus 5.5 and 95.9 percent for GPT-6 Astra, and Moonshot says it can infer shape and proportion from a text brief, sketch or reference image, then script Blender and CAD tools directly. A new 3D Make feature on kimi.ai lets users preview, edit, print or export the models.
Finance is the other showcase. K3.1 Pro claims state-of-the-art 64.4 on OfficeQA Pro v2, a financial-document test combining retrieval, cross-document verification and multi-step numerical reasoning, and its claim-level hallucination rate on Moonshot’s internal hard stress test fell from 23.4 percent for K3 to 8.3 percent. Launch coverage puts the Kimi API price at roughly a third cheaper than K3’s, and independent benchmarker Artificial Analysis scores the model 55 on its Intelligence Index, fourth overall behind only three closed models.
The Cybersecurity Trade-Off
The most unusual part of the launch post is a full section on what Moonshot deliberately took out. As post-training scaled, the company says it observed vulnerability-research capability improving steadily — cyber-related content embedded in the broad training corpus, cyber tasks in the RL environment mix — and its experiments showed that without intervention, downstream cyber capability would rise substantially. For the open-weight release, Moonshot says it identified and removed cyber-related data from both training stages while keeping the general coding tasks the model needs for usefulness.
It published the cost of that choice: on its in-house xdev-bench-hard vulnerability suite, the checkpoint without the risk-reduction measures solves 90.5 percent of problems, while the released K3.1 Pro solves 81.0 percent. Moonshot frames the open-weight decision as a deliberate diffusion of dual-use defensive capability — a model users can run and audit themselves — while pledging not to extend the absolute frontier of cyber capability.
Availability And Open Weights
Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed are live now on Kimi.ai and the Kimi mobile apps, the Kimi Work desktop app (version 3.1.0 or later), Kimi Code via the /model command, and the Kimi API as kimi-k3.1-pro and kimi-k3.1-pro-highspeed, with zero data retention for API customers. Enterprise administrators can enable it, though access is off by default at launch.
The full model weights follow by November 4 — a deliberate two-week window that gives Moonshot a head start on monetizing the hosted version before anyone can self-host. For developers, the release continues the pattern set by recent Chinese open-model drops like DeepSeek’s V4.1-Flash-1017: frontier-adjacent capability arriving first as a cheap API, with the weights to follow.
Why It Matters
K3.1 Pro lands when the gap between the best open-weight models and the closed frontier is measured in single benchmark points rather than whole generations — Moonshot’s own 0.7-point margin behind Claude Opus 5.5 on Terminal-Bench 4.0 is a figure that would have been unthinkable for an open model two years ago. Cheaper inference, a 1-million-token context and open weights within weeks give smaller teams an agent substrate that competes with flagship APIs at a fraction of the cost. And the deliberate cyber-data removal is a template other open-model shops may copy — a published, quantified trade-off between open release and offensive capability, at a time when regulators and the labs are still arguing about where that line sits.
Background
Moonshot AI, the Beijing-based startup behind the Kimi assistant, has built its reputation on aggressively priced, high-performing open-weight releases — Kimi K2 and its K2.5 update were already among the strongest open models on the market. The K3 generation drew the same closed-frontier comparisons that DeepSeek’s 2025 shock produced, and K3.1 Pro is the first attempt to push that lead further. For context, see our explainer on Kimi and Moonshot AI.
Source: Moonshot AI; MadRobot
0 Comments