Kimi K2.6
- Provider
- Moonshot AI
- Status
- available
- Context
- 262,144 tok
- SWE-bench
- 80.2%
- Price
- $0.95 / $4 /MTok
Kimi K2.6 is Moonshot AI’s open-weight flagship, released on 20 April 2026 and built for agentic coding and long-horizon execution. Its standout claim is that it is the first open-weights model to lead SWE-bench Pro, scoring 58.6% — ahead of Claude Opus 4.6 and GPT-5.4 (TokenMix, Verdent). For teams that want frontier-adjacent coding they can run on their own hardware, it is one of the strongest options available.
Architecturally it is a 1-trillion-parameter sparse mixture-of-experts model (32B active per token, 384 experts) with a 256K-token context window, released under a modified MIT licence and shipping natively in INT4 quantisation to keep hosting cheap (Hugging Face, codersera). It also introduces an Agent Swarm mode that scales to hundreds of coordinated sub-agents.
Quick specs
| Provider | Moonshot AI |
| Released | 20 April 2026 |
| Status | Available (open weights) |
| Architecture | Sparse MoE — 1T total / 32B active, 384 experts |
| Context window | 262,144 tokens (256K) |
| Licence | Modified MIT (self-hostable) |
| Modalities | Text, image and video in; text out |
| SWE-bench Pro | 58.6% (first open model to lead it) |
| SWE-bench Verified | ~80.2% |
| Hosted price | ~$0.60–0.95 in / ~$4 out per MTok |
| Best for | Self-hosted agentic coding, long-horizon and multi-agent tasks |
| Limitations | Below the closed frontier on broad capability; coding-focused |
What Kimi K2.6 is
Kimi K2.6 is a native agentic coding model — the focus is reading and writing across real codebases, running multi-step tasks, and staying coherent over long horizons, rather than being a general chatbot (Verdent). It offers Instant, Thinking, Agent and Agent Swarm modes; the Agent Swarm mode is the headline feature, scaling horizontally to as many as 300 sub-agents executing thousands of coordinated steps, decomposing a task into parallel, domain-specialised subtasks (codersera).
Crucially, it is open weights under a modified MIT licence, so all of this can run on your own infrastructure — the appeal for data-sovereign, air-gapped or cost-sensitive teams.
Benchmark performance
Vendor- and third-party-reported figures; treat vendor numbers as a ceiling.
| Benchmark | Kimi K2.6 | Notes |
|---|---|---|
| SWE-bench Pro | 58.6% | First open model to lead it; +5.2pp vs Opus 4.6, +0.9pp vs GPT-5.4 |
| SWE-bench Verified | ~80.2% | Strong, frontier-adjacent on coding |
Kimi K2.6’s profile is coding-first: it leads the open field on agentic software-engineering benchmarks and is competitive with closed flagships there (Verdent). On broad, general-purpose capability it sits below the closed frontier (Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro) and competes with the best open models like DeepSeek V4. See best AI models and best AI for coding for the standings.
Pricing and access
Kimi K2.6 is free to self-host under a modified MIT licence; the weights are on Hugging Face. Hosted pricing is low and varies by provider — roughly $0.60 (Parasail) to $0.95 (Fireworks) per million input tokens and about $4 output (DeepInfra) — and the native INT4 packaging keeps inference cheap. It also powers Moonshot’s Kimi assistant.
How Kimi K2.6 compares
- vs closed flagships — On SWE-bench Pro, Kimi K2.6 actually edges Claude Opus 4.6 and GPT-5.4, a milestone for open weights; on broad capability the closed leaders (Claude Opus 4.8, GPT-5.5) remain ahead.
- vs other open models — Against DeepSeek V4, Alibaba’s Qwen and Llama 4, Kimi K2.6’s edge is agentic coding and the Agent Swarm mode, plus a genuinely permissive (modified MIT) licence.
Known limitations
Coding-focused. Its strength is agentic software engineering; for broad general use the closed flagships are stronger. Below the closed frontier on overall capability. Hosting the full 1T model still needs serious hardware despite INT4 packaging. As always, vendor benchmarks are a ceiling — prefer standardized leaderboards where decisions ride on the number.
FAQ
What is Kimi K2.6?
Kimi K2.6 is Moonshot AI’s open-weight flagship (April 2026): a 1-trillion-parameter sparse mixture-of-experts model focused on agentic coding, with a 256K context and an Agent Swarm mode. It is the first open-weights model to lead SWE-bench Pro.
Is Kimi K2.6 open source?
Yes — it ships as open weights under a modified MIT licence, so you can download, self-host, fine-tune and use it commercially. The weights are on Hugging Face.
How much does Kimi K2.6 cost?
Free to self-host. Hosted via third parties, it runs roughly $0.60–0.95 per million input tokens and about $4 output, depending on the provider.
How good is Kimi K2.6 at coding?
Very strong for an open model — 58.6% on SWE-bench Pro, ahead of Claude Opus 4.6 and GPT-5.4 and the first open-weights model to lead that benchmark, with SWE-bench Verified around 80%.
What is Agent Swarm?
A mode that scales Kimi K2.6 horizontally to as many as 300 sub-agents running thousands of coordinated steps, decomposing a task into parallel, domain-specialised subtasks for complex agentic work.
Last verified 18 June 2026. Kimi K2.6 figures are Moonshot- and third-party-reported (kimi.com, Hugging Face, TokenMix, Verdent, DeepInfra); hosted pricing varies by provider. Confirm against Moonshot’s official pages and your host before relying on specific numbers.