{"slug": "kimi-k3-1-pro-moonshot-s-3t-class-multimodal-model", "title": "Kimi K3.1 Pro: Moonshot's 3T-class multimodal model", "summary": "Moonshot AI released Kimi K3.1 Pro, a 2.8-trillion-parameter open-weight multimodal model it calls \"our most capable 3T-class open-weight model to date,\" with the API live today and full weights promised by Nov 4. On Moonshot's published benchmarks, Kimi K3.1 Pro scored 65.7 on Terminal-Bench 4.0, behind Claude Opus 5.5 (66.4) but ahead of GPT-6 Astra (58.2) and Fable 5.1 (57.9), and took first place on SWE Marathon at 51.7, ten points over Opus 5.5 (45.3). Moonshot said the open-weight release deliberately removes cyber-related data from both mid-training and RL, stating \"we aim not to release open weights that would extend the absolute frontier of cyber capability.", "body_md": "Moonshot AI’s own line about Kimi K3.1 Pro: “our most capable 3T-class open-weight model to date.” The announcement landed about two hours ago and the claims worth your attention are not the 2.8-trillion-parameter headline - it’s the same MoE shape as Kimi K3 from September - but what the post-training now does: full-project game coding across three engines, parametric 3D modeling in Blender and CAD by writing scripts without connectors, and complete financial workflows delivered end to end.\n\n## The numbers from the official blog\n\nMoonshot published a tech blog alongside the announcement with the comparison panels. The honest reading:\n\n- \n**Terminal-Bench 4.0 (coding agents on realistic terminal tasks): K3.1 Pro 65.7** , ahead of GPT-6 Astra (58.2) and Fable 5.1 (57.9), a hair behind Claude Opus 5.5 (66.4). The board’s footnote says every model was maxed out on thinking effort, so the comparison is at full strength on all sides.\n- \n**SWE Marathon: 51.7, first place** , ten points over Opus 5.5 (45.3) and Fable 5.1 (45.6). Long-horizon software work is where the gains concentrated.\n- \n**BenchCAD with tools: 92.7** - reconstructing 3D parts from multi-view renderings by writing CAD code - approaching Opus 5.5 (96.2) and GPT-6 Astra (95.9). This is the parametric-3D claim made concrete.\n- \n**OfficeQA Pro v2: 64.4** (best on the board),**OSWorld V2.1: 79.6** ,**Humanity’s Last Exam Full w/ tools: 70.0** (above Opus 5.5’s 67.7),**GDP.pdf 34.6** ,**ZeroBench w/ tools pass@5: 63.0** .\n- The charts admit their holes: Terminal-Bench Science 0.1 goes to GPT-6 Astra (68.1) with K3.1 Pro at 38.1 - a real weakness, printed in their own panel.\n\n## The 3D and game-engine story\n\nThe differentiators beyond raw agentic coding: K3.1 Pro builds complete experiences in Unity, Unreal Engine, and Godot - combining engine calls with web search and 3D generation - and its BenchCAD numbers back the spatial-reasoning claim. Without MCP connectors it writes and runs scripts for tools like Blender and CAD software to build and refine assets, edit components, adjust geometry and materials. Moonshot also exposes the outputs as products: preview and edit 3D models online, download print-ready files, export as video or Live Photos.\n\n## The two release facts that matter\n\nFirst: **the model is live today on the API, the weights are not.** Moonshot promises the full open weights by **Nov 4**. Until then this is a platform-only frontier model - day-zero hosting stories will follow the weights drop.\n\nSecond, and this is the unusual part: Moonshot says the open-weight release deliberately **removes cyber-related data from both training stages** (mid-training and RL) while keeping the general-purpose coding work. Their framing: “we aim not to release open weights that would extend the absolute frontier of cyber capability.” It is a published, bounded safety decision on an open release, which is rare anywhere in the open-weights world.\n\n## What it means for the local crowd\n\nThe K3.1 Pro shape repeats the K3 lesson: 2.8T total means frontier-class quality, 104B active means serving speed follows what a 104B model reads per token. No sane home cluster runs 2.8T at 4-bit (1.4 TB) - this stays a rent-the-FP8-or-quantized-inference-weights model, or wait for the usual community quants to make it a pooled-node project. The platform versions (kimi-k3.1-pro and the 2-3x-highspeed kimi-k3.1-pro-highspeed) are the practical way to judge the model this week.\n\nSources: Moonshot’s tech blog (kimi.ai/blog/kimi-k3-1-pro), the announcement post (x.com/Kimi_Moonshot), benchmark figures as published by Moonshot AI.\n\nRelated on this site: the deepseek-v4-1-flash rows and the engine catalog are where the open-weights ladder lives; the models index tracks the whole family as rows.", "url": "https://wpnews.pro/news/kimi-k3-1-pro-moonshot-s-3t-class-multimodal-model", "canonical_source": "https://tokenstead.ai/guides/kimi-k31-pro", "published_at": "2026-10-23 19:41:46+00:00", "updated_at": "2026-10-23 19:43:04.518349+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-safety", "ai-research", "ai-products"], "entities": ["Moonshot AI", "Kimi K3.1 Pro", "Kimi K3", "Claude Opus 5.5", "GPT-6 Astra", "Fable 5.1", "Unity", "Blender"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/kimi-k3-1-pro-moonshot-s-3t-class-multimodal-model", "markdown": "https://wpnews.pro/news/kimi-k3-1-pro-moonshot-s-3t-class-multimodal-model.md", "text": "https://wpnews.pro/news/kimi-k3-1-pro-moonshot-s-3t-class-multimodal-model.txt", "jsonld": "https://wpnews.pro/news/kimi-k3-1-pro-moonshot-s-3t-class-multimodal-model.jsonld"}}