Kimi K2-Instruct-0905

// Overview

Kimi K2-Instruct-0905 is an updated checkpoint of Moonshot AI's Kimi K2 model line, released on September 9, 2025. Like the original Kimi K2-Instruct, it is a Mixture-of-Experts (MoE) language model with 1 trillion total parameters but only 32 billion activated per token, which keeps inference costs closer to a mid-sized dense model while drawing on the capacity of a much larger one. The weights are published openly under a Modified MIT License on Hugging Face.

The 0905 update focuses on three things: stronger agentic coding (higher accuracy and better generalization across different coding scaffolds and harnesses), a more polished frontend-coding experience (more aesthetic and functional output for web, 3D, and UI tasks), and an extended context window, raised from 128K to 256K tokens to better handle multi-file edits, large-repository question answering, and long-horizon agent runs.

Architecturally it uses 384 experts with 8 routed plus 1 shared expert selected per token, 61 layers, 64 attention heads with MLA attention, and SwiGLU activations. It is a text-only, non-reasoning instruct model, positioned for tool use and coding agents rather than long chain-of-thought reasoning. It can be self-hosted on vLLM, SGLang, KTransformers, or TensorRT-LLM, or accessed through Moonshot's OpenAI/Anthropic-compatible API and several third-party inference providers.

// Benchmarks

1. SWE-bench Verified (agentic, single attempt)69.2%
2. SWE-bench Multilingual (agentic, single attempt)55.9%
3. Multi-SWE-Bench33.5%
4. Terminal-Bench44.5%
5. SWE-Dev66.6%
6. LiveCodeBench v6 (Pass@1)53.7%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Representative listed price; open weights can also be self-hosted, and rates vary across inference providers.

Pricing source ↗

// Strengths

• Agentic coding: improved accuracy and generalization across coding scaffolds, with strong SWE-bench Verified and SWE-Dev scores
• Frontend and UI generation: more aesthetic and functional output for web, 3D, and interface tasks
• 256K-token context for multi-file edits, large-repo Q&A, and long-horizon agent tasks
• Open weights under a permissive Modified MIT License, allowing self-hosting and commercial use
• Efficient MoE design: trillion-parameter capacity with only 32B parameters active per token
• Strong tool-use and function-calling behavior for building autonomous agents

// Best for

• Autonomous coding agents that edit, build, and test across multi-file repositories
• Frontend and web UI generation, including 3D and interactive layouts
• Repository-scale question answering and code understanding over large codebases
• Tool-calling and function-calling workflows for agent frameworks
• Self-hosted open-weight deployments where data control or cost predictability matters
• Long-context document and codebase processing up to 256K tokens

// How to access

// Kimi K2 — every version

The full lineage of the Kimi K2 line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.