Kimi K3.1 Pro launches with 1M-token context, 65.7% on Terminal-Bench 4.0; open weights due by Nov 4

Kimi K3.1 Pro, the first model in the new K3.1 series, scores 65.7% on Terminal-Bench 4.0 and ships with a 1-million-token context window. Kimi describes it as its most capable 3T-class open-weight model to date. It is available today on Kimi.com, Kimi Work, Kimi Code and the Kimi API. The full model weights will be released by Nov 4.

Kimi also released Kimi K3.1 Pro Highspeed, which it says delivers 2-3× the speed of K3.1 Pro. On the Kimi API, K3.1 Pro costs $2.00 per million input tokens and $10.00 per million output tokens ($0.20 for cached input), and K3.1 Pro Highspeed costs $4.00 and $20.00, according to Moonshot's pricing page on October 23. Kimi K3 is listed at $3.00 and $15.00.

What changed

K3.1 Pro is a post-trained version of Kimi K3 with substantially larger post-training compute. According to Kimi, it optimized the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling and a shared environment quality-check platform to support joint RL across hundreds of downstream domains.

Benchmarks

All figures below are reported by Kimi.

• Terminal-Bench 4.0: K3.1 Pro scores 65.7%, ahead of GPT-6 Astra (58.2%) and narrowly behind Claude Opus 5.5 (66.4%).
• Kimi Code Bench 3.0 Long-horizon (internal): 61.2%, versus 55.95% for K3.
• BenchCAD (with tools): 92.7%, versus Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%).
• OfficeQA Pro v2: 64.4, which Kimi describes as state-of-the-art.
• Internal finance benchmarks: Financial Modeling improves 7.2 points over K3, Financial Research improves 11.6 points, and Business Operation reaches about 4.2× K3's performance. Kimi also reports improvement over K3 on Frontier Finance.

Coding and 3D

Kimi says K3.1 Pro works across Unity, Unreal Engine and Godot, combining them with web search and 3D generation. In an internal long-horizon test, it 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, the simulated human went from collapsing within seconds to walking continuously for 11.2 seconds. Kimi says the model identified and corrected 36 errors in shoulder-muscle moment-arm calculations after reward shaping failed to improve arm motion.

For 3D work, Kimi says the model can write and run scripts for tools such as Blender and CAD software without MCP connectors, and is strong at parametric modeling. A 3D Make feature is now live on kimi.com and the Kimi app, with preview, editing, 3D-print-ready download, and export to video or Live Photos.

Factual accuracy

On Kimi Hallu Bench Hard, an internal stress test, claim-level hallucination rate fell from 23.4% for K3 to 8.3%. The share of responses with at least one factual error was 42.4% of its predecessor's rate. Kimi notes these rates do not represent overall error rates in daily use.

Cybersecurity controls

Kimi says vulnerability research performance rose steadily as post-training scaled, and traced this to cyber-related content in the general mid-training corpus and cyber-related tasks in the RL environment mix. For the open-weight release, it removed cyber-related data from both training stages while retaining general-purpose coding tasks. Kimi says that without removal at either stage, downstream cyber capability would increase substantially, based on preliminary experiments on internal models. It evaluated the cleaned model on xdev-bench-hard, an in-house benchmark of source-code-based vulnerability discovery and exploitation tasks.

What this means

Kimi positions K3.1 Pro as an open-weight model that matches or approaches proprietary systems on coding, agentic and knowledge benchmarks, though on the figures it reports, Claude Opus 5.5 still leads on Terminal-Bench 4.0 and BenchCAD. Most headline results come from Kimi's own evaluations, and the internal benchmarks (Kimi Code Bench, Hallu Bench Hard, finance suites) cannot be independently checked until the weights ship by Nov 4. The data-removal approach to limiting cyber capability is notable for an open-weight release, since controls applied during training cannot be reversed by API-level policy once weights are public.