Moonshot AI's 2026-04 flagship — 1T-parameter MoE that ties GPT-5.5 on coding benchmarks. Agent swarm scales to 300 sub-agents and 4000 coordinated steps. Open-weight.
Kimi K2.6, released April 20, 2026 by Moonshot AI, is a 1-trillion-parameter, 32-billion-active Mixture-of-Experts (MoE) model purpose-built for long-horizon coding, agent swarms, and multimodal input. Its architecture deploys 384 experts (8 routed + 1 shared per token) across 61 layers with Multi-head Latent Attention (MLA), SwiGLU activation, a 160K-token vocabulary, native INT4 quantization, and a 400M-parameter MoonViT vision encoder. The model is released as open weights under a Modified MIT License.
K2.6 is the latest flagship in the Kimi lineage, succeeding K2.5 with a major focus on agentic reliability at scale: it supports swarms of up to 300 parallel sub-agents executing 4,000 coordinated steps, sustained autonomous runs of 12+ hours, and a new Claw Groups framework for multi-agent/human collaboration.
| Type | Rate |
|---|---|
| Input | $1.03 / 1M tokens |
| Output | $4.32 / 1M tokens |
| Release date | 2026-04-20www.kimi.com ↗ | verified |
| Architecture | 1T-total / 32B-active MoE; 384 experts; 61 layers; MLA; SwiGLU; native INT4www.kimi.com ↗ | verified |
| Context window | 256,000 tokenswww.kimi.com ↗ | verified |
| License | Modified MIT Licensehuggingface.co ↗ | verified |
| Training data cutoff | Not publicly disclosed | unknown |
| Agent swarm scale | Up to 300 sub-agents, 4,000 coordinated stepswww.kimi.com ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
SWE-Bench Pro Ties GPT-5.4 (57.7), exceeds Opus 4.6. | 58.6% | kimi.com/blog ↗ | |
SWE-Bench Verified | 80.2% | MarkTechPost ↗ | |
Terminal-Bench 2.0 | 66.7% | MarkTechPost ↗ | |
LiveCodeBench v6 Pass@1. | 89.6% | MarkTechPost ↗ | |
HLE (Full, w/ tools) Leads GPT-5.4 (52.1), Opus 4.6 (53.0). | 54.0% | MarkTechPost ↗ | |
BrowseComp (Agent Swarm) | 86.3% | MarkTechPost ↗ | |
DeepSearchQA | 92.5f1 | MarkTechPost ↗ | |
Toolathlon | 50.0% | Latent Space ↗ | |
MathVision w/ python | 93.2% | Latent Space ↗ | |
CharXiv w/ python | 86.7% | Latent Space ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "moonshot/kimi-k2.6",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Standard OpenAI chat completions.
curl https://api.therouter.ai/v1/chat/completions -H "Authorization: Bearer $THEROUTER_API_KEY" -H "Content-Type: application/json" -d '{"model":"moonshot/kimi-k2.6","messages":[{"role":"user","content":"Hello"}]}'Moonshot AI open-sourced Kimi K2.6 under Modified MIT License. 1T-parameter MoE with 300-agent swarms, 4,000-step execution, native INT4, Claw Groups. Day-0 in vLLM, SGLang, KTransformers, MLX, Ollama, OpenRouter.
Yes. Modified MIT License, FP16 and INT4 weights on Hugging Face (moonshotai/Kimi-K2.6). Runs on vLLM, SGLang, KTransformers, MLX, Ollama. Requires transformers >=4.57.1, <5.0.0.
K2.6 ties or beats both on agentic benchmarks: SWE-Bench Pro 58.6 (Opus 4.6: 53.4), HLE w/tools 54.0 (GPT-5.4: 52.1), LiveCodeBench v6 89.6 (Opus 4.6: 88.8), at ~80% lower cost.
Yes. Native reasoning enabled by default, returns thinking via reasoning_content. Set include_thoughts=false to skip.
K2.6 decomposes tasks into heterogeneous subtasks executed by self-created domain-specialized sub-agents. Scales to 300 parallel sub-agents and 4,000 coordinated steps. Supports Claw Groups.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | www.kimi.com ↗ | 2026-05-28 | verified |
| Architecture | www.kimi.com ↗ | 2026-05-28 | verified |
| Context window | www.kimi.com ↗ | 2026-05-28 | verified |
| License | huggingface.co ↗ | 2026-05-28 | verified |
| Training data cutoff | — | — | unknown |
| Agent swarm scale | www.kimi.com ↗ | 2026-05-28 | verified |
| SWE-Bench Pro | kimi.com/blog ↗ | 2026-05-28 | verified |
| SWE-Bench Verified | MarkTechPost ↗ | 2026-05-28 | verified |
| Terminal-Bench 2.0 | MarkTechPost ↗ | 2026-05-28 | verified |
| LiveCodeBench v6 | MarkTechPost ↗ | 2026-05-28 | verified |
| HLE (Full, w/ tools) | MarkTechPost ↗ | 2026-05-28 | verified |
| BrowseComp (Agent Swarm) | MarkTechPost ↗ | 2026-05-28 | verified |
| DeepSearchQA | MarkTechPost ↗ | 2026-05-28 | verified |
| Toolathlon | Latent Space ↗ | 2026-05-28 | verified |
| MathVision w/ python | Latent Space ↗ | 2026-05-28 | verified |
| CharXiv w/ python | Latent Space ↗ | 2026-05-28 | verified |
| Kimi K2.6 released: open-weight 1T MoE ties GPT-5.5 on SWE-Bench Pro | kimi.com/blog ↗ | 2026-05-28 | verified |
| Is Kimi K2.6 open-weight? | huggingface.co ↗ | 2026-05-28 | to verify |
| How does K2.6 compare to Opus 4.6 and GPT-5.5? | www.marktechpost.com ↗ | 2026-05-28 | to verify |
| Does K2.6 support reasoning? | openrouter.ai ↗ | 2026-05-28 | to verify |
| What is the Agent Swarm feature? | www.kimi.com ↗ | 2026-05-28 | to verify |