Kimi drops K3.1 Pro with 1M context and 3D coding tools

Most coding assistants are great at fixing small bugs or knocking out a simple web front-end. But ask one to build a whole 3D interactive experience across game engines or solve structural dynamics for hundreds of human muscles, and they usually crumble.

Kimi officially launched K3.1 Pro, its most capable 3T-class open-weight MoE model, alongside a 2–3× faster Highspeed variant. The multimodal heavyweight packs 2.8T total parameters, 104B active parameters, and a massive 1M-token context window. While open weights will not land until Nov 4, 2026, the model is already making serious waves across the industry.

Long-horizon coding meets parametric 3D

The real headline with K3.1 Pro is how it tackles multi-step, hours-long engineering challenges.

In one internal stress test, the model built an interactive walking simulator from scratch using human-atlas, an open-source biological framework with 296 skeletal parts and 402 muscle components. Across 167 tasks and 420 reviews, K3.1 Pro did not just tweak reinforcement learning rewards blindly when arm movements failed. Instead, it ran open-loop physical probes, discovered 36 faulty moment-arm calculations in the shoulder muscles, corrected the math, and retrained the controller until the simulated human walked continuously for 11.2 seconds.

For 3D development, it writes full-scale code for Unreal Engine, Unity, and Godot. On BenchCAD, which evaluates CAD code generation from multi-view images, K3.1 Pro scored 92.7% with tools, closing the gap with Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%).

Impressively, it controls Blender and CAD software directly via raw scripts without needing MCP connectors. Kimi also pushed a new 3D Make feature directly to its web app and mobile client, letting users preview parametric models, export them for 3D printing, or render Live Photos.

Beating Astra at a fraction of the cost

K3.1 Pro is also built for deep document intelligence. Moonshot claims the model reduces financial hallucinations by 65%, hitting a state-of-the-art 64.4 on OfficeQA Pro v2.

On Terminal-Bench 4.0, which tests coding agents on rigorous, resource-capped command-line environments, K3.1 Pro put up a 65.7% score. That easily tops GPT-6 Astra (58.2%) and sits just behind Claude Opus 5.5 (66.4%).

Independent testing from Artificial Analysis ranked K3.1 Pro fourth overall on its Intelligence Index with a score of 55, beating GPT-6 Astra and landing just behind Anthropic's top models and Meta's Muse Wonder. It also emerged as the top openly available cyber-defense model with a 64% score, trailing only a specialized GPT-6 variant strictly reserved for vetted defenders.

On the API side, K3.1 Pro costs $2.00 per million input tokens and $10.00 per million output tokens, roughly one-fifth the cost of GPT-6 Astra and on par with Claude Sonnet 5.5. The catch? It runs a bit slow at 73.5 tokens per second, and Artificial Analysis noted it is extremely verbose, burning 150 million tokens across its evaluation suite.

Why it matters

For builders and creators, K3.1 Pro shifts the role of AI from a smart autocomplete into an actual junior 3D asset technician. If you build mechanical parts in CAD, script visual assets in Blender, or prototype indie games in Godot, this model is worth testing immediately. Just do not expect to run it on your home PC when open weights drop on Nov 4, 2026: a 2.8T MoE monster is strictly reserved for high-end enterprise clusters, meaning cloud endpoints will remain the standard.

For the frontier AI landscape, this launch should ring alarm bells in Silicon Valley. Open models are no longer lagging a generation behind in specialized domains. By matching or beating GPT-6 Astra on hardcore engineering tasks and cyber defense, Moonshot proves that aggressive RL post-training can turn massive parameter scales into genuine reasoning muscle.

We think Moonshot played this very smartly by focusing on hard-to-fake, tool-integrated workflows like mechanical CAD and sandbox code execution rather than chasing generic chat vibes. The main headache for users right now will be token consumption and latency: K3.1 Pro loves to write extensive answers, and even with the Highspeed version, running multi-turn loops at production scale will require careful prompt management to keep costs and response times in check.