Skip to content

Kimi K2.6

Modified MIT

Moonshot AI · 1000B · transformer-moe

2026-04-20262K context1000B params

Use Cases

chatcodereasoningmultilingualvisiontools

Quantization Options

QuantBitsVRAMQualityStatus
Q2_Krec2360.0 GBModerate
Q4_K_M4604.0 GBGood

About this model

Kimi K2.6 is Moonshot AI's natively multimodal agentic flagship — a 1-trillion-parameter Mixture-of-Experts with 32B active per token and 256K context, released under a Modified MIT license in April 2026. Built for long-horizon coding and agent-swarm orchestration (Moonshot demonstrated 13-hour continuous coding sessions and 300-agent clusters), it scores 80.2 on SWE-bench Verified and ties frontier closed models on SWE-bench Pro at a fraction of the API cost. Locally this is datacenter-class hardware only: the most aggressive 2-bit community quant is a 340 GB download needing ~360 GB of combined memory. The official Ollama library carries only the `kimi-k2.6:cloud` remote tag, which is how nearly everyone actually runs it.

Benchmarks

80.2
swe-bench-verified
58.6
swe-bench-pro