Kimi K2.6

// Overview

Kimi K2.6 is Moonshot AI's flagship open-weight model in the Kimi K2 line, released on April 20, 2026. It is a 1-trillion-parameter Mixture-of-Experts (MoE) model that activates 32 billion parameters per token, extending the K2 family's focus on autonomous, long-horizon agentic coding. The weights are published on Hugging Face (moonshotai/Kimi-K2.6) under a Modified MIT license, and the model is served on Kimi.com, the Kimi app, the official Moonshot API, and the Kimi Code CLI.

Architecturally, Kimi K2.6 uses 384 experts (8 routed plus 1 shared per token) with Multi-head Latent Attention to keep the KV-cache footprint small, and folds in a 400M-parameter MoonViT encoder so the single model natively reads images and video alongside text. It supports a 256K-token (262,144) context window and ships in native INT4 quantization produced via Quantization-Aware Training, keeping the ~594GB checkpoint deployable on standard inference stacks like vLLM, SGLang, and KTransformers.

The release leans hard into agentic workflows: K2.6 offers Instant, Thinking, Agent, and Agent Swarm variants, with the swarm mode scaling to 300 domain-specialized sub-agents executing up to 4,000 coordinated steps in a single autonomous run. On coding and agentic benchmarks it trades blows with the leading proprietary frontier models of its day while remaining fully open-weight, which is the core of its appeal.

// Benchmarks

1. SWE-Bench Verified80.2%
2. SWE-Bench Pro58.6%
3. Humanity's Last Exam (HLE-Full, with tools)54%
4. LiveCodeBench v689.6%
5. AIME 202696.4%
6. GPQA-Diamond90.5%
7. BrowseComp (Agent Swarm mode)86.3%
8. MathVision (with Python)93.2%

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

// Pricing

Official Moonshot API pricing. Third-party providers (OpenRouter, DeepInfra, Fireworks) list their own rates.

Pricing source ↗

// Strengths

• State-of-the-art open-weight performance on agentic coding benchmarks (e.g. SWE-Bench Pro 58.6, SWE-Bench Verified 80.2)
• Fully open weights under a permissive Modified MIT license, self-hostable on vLLM/SGLang/KTransformers
• Native multimodal input (text, image, video) via the built-in MoonViT encoder
• Agent Swarm orchestration scaling to 300 sub-agents and 4,000 coordinated steps for long-horizon autonomous tasks
• 256K-token context window for large codebases and documents
• Aggressive pricing with an ~83% discount on cached input tokens

// Best for

• Autonomous, multi-step software engineering and bug-fixing over large repositories
• Multi-agent orchestration and long-horizon research/coding runs
• Self-hosted deployment where open weights and data control matter
• Deep web research and agentic search workflows
• Coding-driven UI/UX generation from prompts and visual inputs
• Math and reasoning tasks with thinking mode enabled

// 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.