Kimi K2.6 — Open-Source Model for Agentic Coding

Trusted by users from 10,000+ companies

Core Capabilities You Should Explore

Kimi K2.6 comes packed with capabilities that add up across every kind of technical and knowledge work.

Native Multimodal

Accepts text, images, and video natively. Analyse charts, screenshots, and footage without a separate pipeline.

256K Context

Holds entire codebases, long documents, or extended conversation histories without losing track of earlier content.

Thinking Mode

Built-in long-thinking capability for complex reasoning. Toggle it on for hard problems, off for speed-sensitive tasks.

Frequently Asked Questions

Quick answers covering everything you need to know about this new model.

Reusable Skills

Upload a high-quality document once and K2.6 captures its structure and style as a reusable skill for future outputs.

Proactive Agents

Runs as a persistent background agent for days, managing schedules, monitoring systems, and resolving incidents without human prompting.

OpenAI Compatibility

Drop-in compatible with the OpenAI SDK. Integrate K2.6 into any existing stack without rewriting your API calls.

Open Weights

Model weights are publicly available on HuggingFace. Run it locally, fine-tune it, or deploy it on your own infrastructure.

Cross-language Coding

Reliable generalisation across Rust, Go, Python, and front-end stacks, not just the languages it was most trained on.

Bring Your Agents

Claw Groups lets external agents running any model on any device join a shared K2.6-coordinated swarm as collaborators.

Proven Performance Across the Toughest AI Benchmarks

Kimi K2.6 competes directly with the best closed-source models across benchmarks that reflect real engineering work.

Leader in Multi-source Research Tasks

At 92.5 F1, Kimi K2.6 answers complex, source-heavy research questions with high accuracy, making it a reliable choice for agentic information work.

Ahead of Competitors in Coding

With 58.6% on SWE-Bench Pro, K2.6 resolves more than half of real-world production code issues autonomously, placing it among the top-ranked models globally on this evaluation.

Autonomous Computer-use Completion

Kimi K2.6 completes nearly three quarters of fully autonomous computer-use tasks correctly, a 10-point improvement over its predecessor K2.5.

End-to-end Engineering Execution

K2.6 handles complex coding projects autonomously across thousands of tool calls, spanning hours of continuous work in different coding languages. It iterates, optimises, and delivers.

Parallel Agent Teams on a Single Prompt

K2.6's swarm architecture deploys up to 300 specialised sub-agents simultaneously, each handling a different part of the task. A single prompt can produce a research report, a spreadsheet, and a website, all in one autonomous run.

Production Interfaces from a Description

Kimi K2.6 builds complete, visually considered front-end interfaces from a single description, including scroll animations, authentication flows, and database integrations. It delivers something you can actually ship.

Kimi K2.6: Open-Source AI for Long-Horizon Work

Core Capabilities You Should Explore

Native Multimodal

256K Context

Thinking Mode

Frequently Asked Questions

What is Kimi K2.6?

How much does Kimi K2.6 cost via the API?

Is Kimi K2.6 open source?

What context window does Kimi K2.6 support?

Does Kimi K2.6 support thinking mode?

How does K2.6 compare to K2.5?

Can I use Kimi K2.6 with an existing OpenAI SDK setup?

What is Kimi's agent swarm and what can it do?

Where can I access Kimi K2.6?