About the news
Positioned as a major step forward for open intelligence, the model is a heavily post-trained version of Kimi K3 that received substantially more reinforcement-learning compute across hundreds of domains. It is already live across Kimi’s products and API, with full open weights promised by November 4.
Key features:
- Architecture & scale: Mixture-of-Experts (MoE) model with approximately 2.8 trillion total parameters and 104 billion active parameters per token; multimodal (text + vision); 1-million-token context window.
- Speed variant: Kimi K3.1 Pro Highspeed delivers the same capabilities at roughly 2–3× the speed of the standard Pro model.
- Availability & pricing: Live now on kimi.ai, mobile apps, Kimi Work, Kimi Code, and the Kimi API (model IDs: kimi-k3.1-pro and kimi-k3.1-pro-highspeed). API pricing starts at $2 / $10 per million input/output tokens for the standard model (with lower cache-hit rates); Highspeed is higher. Full open weights scheduled for release by November 4, 2026.
- Other notes: Improved reliability for finance and business-operation simulation; designed for agentic workflows and long-horizon tasks.
Performance highlights (company-reported, mostly at max reasoning effort):
- Terminal-Bench 4.0: 65.7% (narrowly behind Claude Opus 5.5 at 66.4%, ahead of GPT-6 Astra at 58.2%).
- SWE-Marathon: 51.7% (leads the listed comparisons).
- BenchCAD (3D/CAD reconstruction with tools): 92.7%.
- Strong gains in long-horizon coding, full-project game development (Unreal, Unity, Godot), parametric 3D modeling in Blender and CAD tools, financial modeling/research, and agentic tasks.
- Hallucination rate on hard internal factual tests cut significantly (claim-level rate down from 23.4% on K3 to 8.3%).
Why it matters
Kimi K3.1 Pro narrows the gap between open-weight and proprietary frontier models on coding and agentic benchmarks while offering a practical 1M-token context and forthcoming open weights. With the Highspeed option for interactive use and broad product integration already in place, it gives developers and enterprises a competitive new choice for complex, multi-step work ahead of the November weight release.
