Kimi K2.6
Moonshot AI·
LLMsopen-weightFrontier
Overview
1T MoE (32B active) open-weight successor to K2.5 with native multimodal vision (MoonViT 400M), 256K context, and agent swarm support for up to 300 sub-agents across 4,000 coordinated steps.
Capabilities and innovations
1T total / 32B active MoE (384 experts, 8 routed + 1 shared)256K context windowNatively multimodal (MoonViT 400M — text, image, video)Agent swarms (300 sub-agents, 4,000 coordinated steps)Long-horizon coding (Rust, Go, Python)Modified MIT licenseMLA (Multi-head Latent Attention)300 Sub-Agent Swarm (up from 100 in K2.5)4,000 Coordinated Tool-Call StepsNative Video Input
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| Humanity’s Last Exam (no tools) | 36.4% | vendor HLE 36.4 no-tools (text-only), recorded for cross-model comparability. Moonshot also reports 54.0 with tools (HLE-Full) and 55.5 with-tools text-only, which are not comparable to the no-tools HLE used here. Corrected from a previously stored with-tools value (54.0). |
| SWE-bench Verified | 80.2% | unsourced |
| SWE-bench Pro | 58.6% | unsourced |
| Terminal-Bench 2.x | 66.7% | unsourced |
| MMMU-Pro | 79.4% | unsourced |
Architecture and hardware
- Parameters
- 1000B total, 32B active per token (MoE)
- Estimated VRAM at Q4
- ~575 GB, Frontier class
- Quantization formats
- INT4, BF16, GGUF
- Recommended runtime
- vLLM
- License
- Modified MIT
Links
More from Moonshot AI
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.