Kimi K3.1 Pro is Moonshot AI’s new open-weight model that the Chinese company says comes close to leading proprietary systems on coding, agent work, and knowledge tasks. In Moonshot’s own figures, it scores 65.7% on Terminal-Bench 4.0—ahead of OpenAI’s GPT-6 Astra at 58.2% and just behind Anthropic’s Claude Opus 5.5 at 66.4%. For readers comparing frontier tools, that is the core claim: a cheaper open-weight option that, on several company tests, outruns GPT-6 Astra and sits near Claude.
How the benchmarks stack up
Moonshot places Kimi K3.1 Pro as the first release in its K3.1 series and its strongest 3T-class open-weight model so far. It is not a new base design: the firm describes it as a post-trained edition of Kimi K3, the 2.8-trillion-parameter model launched in July, trained with far more reinforcement-learning compute across hundreds of domains after upgrades to its RL stack, long-context inference, sandboxes, and rollout scheduling.
All comparisons Moonshot published are self-reported, with K3.1 Pro at maximum reasoning effort, against Claude Opus 5.5, Claude Fable 5.1, GPT-6 Astra, DeepSeek-V4.1-Flash-1017, and GLM-5.3. The company says K3.1 Pro leads SWE-Marathon (51.7 versus 46.3 for Opus 5.5) and Humanity’s Last Exam Full with tools (70.0 versus 67.7). It also claims tops on OfficeQA Pro v2 (64.4), Agents’ Last Exam (43.2), GDP.pdf (34.6), ZeroBench with tools (63.0), BrowseComp (92.1), AutomationBench-AA (73.3), and DeepSWE v1.1 among the proprietary set. On GDPval-AA v2.1 it reports an Elo of 1,846, edging GPT-6 Astra while trailing Opus 5.5 at 1,866.
Gaps remain. On Terminal-Bench Science 0.1 it scores 38.1, well behind GPT-6 Astra (68.1) and Opus 5.5 (63.3). It trails on ProgramBench (31.5 versus 65.0 for Opus), HLE-Diamond (39.4 versus 65.8 for GPT-6 Astra), and GPQA-Diamond (94.2 versus 96.1). Moonshot says weaker cybersecurity results on ExploitBench and ExploitGym are intentional. Against other open models on Terminal-Bench 4.0, DeepSeek-V4.1-Flash-1017 at 42.3 and GLM-5.3 at 41.8 sit more than 20 points behind.
Coding, 3D work, and availability
Moonshot says the model can handle quick codebase fixes through full project builds, including work with Unity, Unreal Engine, and Godot alongside web search and 3D generation. In one internal demo, it built a walking simulator on an open anatomical model with 296 skeletal and 402 muscle parts; over 167 missions the simulated figure improved from collapsing in seconds to walking 11.2 seconds continuously, after the system found and fixed 36 shoulder-muscle calculation errors. On BenchCAD with tools it scores 92.7%, near Opus 5.5 (96.2%) and GPT-6 Astra (95.9%); without tools, 72.8 versus 73.0 for Opus. The company also says it can turn a text brief, sketch, or reference image into a usable 3D model and drive Blender and CAD scripts.
Kimi K3.1 Pro Highspeed is offered as a two-to-three-times faster sibling. Both versions are available now on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, and Moonshot says API use is cheaper than the predecessor. Full open weights are scheduled for release by 4 November. Since the 23 October launch, social feeds have filled with 3D demos, games, and comparison clips—though some observers have asked whether the model itself is producing those assets or relying on external pipelines.
Timeline: what has happened so far
- July 2026: Moonshot launches Kimi K3, a 2.8-trillion-parameter model that later becomes the base for K3.1 Pro post-training.
- 23 October 2026: Kimi K3.1 Pro launches; demos and comparison videos circulate widely online.
- 24 October 2026 onward: Moonshot publicises benchmark tables placing the model ahead of GPT-6 Astra on Terminal-Bench 4.0 and near Claude Opus 5.5 on several agent and coding suites.
- By 4 November 2026: Full model weights are due for open release, according to Moonshot.
Aasan alfaaz mein
Moonshot AI ne naya open model Kimi K3.1 Pro launch kiya hai. Company ke mutabiq ye coding aur agent tests mein GPT-6 Astra se aage aur Claude Opus 5.5 ke qareeb hai. Abhi Kimi.ai aur API par mil sakta hai, aur weights 4 November tak open hone hain. Pakistani developers ke liye matlab ye hai ke ek sasta, open-weight option frontier models ke qareeb claim kar raha hai—lekin scores company ke apne hain, is liye apne kaam par khud check karna behtar hai.
