Moonshot AI Open-Sources Kimi K2.6: Leads HLE-Full and SWE-Bench Pro Over Every Frontier Proprietary Model
Moonshot AI has open-sourced Kimi K2.6, a 1-trillion-parameter mixture-of-experts model with 32 billion active parameters per token, and it leads every frontier proprietary model on two of the most demanding agentic benchmarks available today.
The Numbers That Matter
On HLE-Full with tools — Humanity’s Last Exam in its hardest, tool-augmented form — K2.6 scores 54.0. The comparison:
| Model | HLE-Full (tools) |
|---|---|
| Kimi K2.6 | 54.0 |
| Claude Opus 4.6 | 53.0 |
| GPT-5.4 | 52.1 |
| Gemini 3.1 Pro | 51.4 |
On SWE-Bench Pro, which tests resolution of real-world professional GitHub issues:
| Model | SWE-Bench Pro |
|---|---|
| Kimi K2.6 | 58.6 |
| GPT-5.4 (xhigh) | 57.7 |
| Gemini 3.1 Pro (thinking) | 54.2 |
| Claude Opus 4.6 (max effort) | 53.4 |
| Kimi K2.5 | 50.7 |
SWE-Bench Verified comes in at 80.2% — within a point of the current frontier pack. LiveCodeBench v6 scores 89.6, narrowly ahead of Claude Opus 4.6 at 88.8. On Terminal-Bench 2.0 using Terminus-2 agent, K2.6 hits 66.7, behind Gemini 3.1 Pro (68.5) but ahead of GPT-5.4 and Claude Opus 4.6 (both 65.4).
BrowseComp in Agent Swarm mode posts 86.3, up from 78.4 on K2.5.
Architecture
K2.6 runs 384 experts — 8 routed plus 1 shared per token — with multi-head latent attention (MLA), 256K context window, native multimodality, and INT4 quantisation. Weights are on Hugging Face under a Modified MIT License with day-zero support in vLLM, OpenRouter, Cloudflare Workers AI, Baseten, and MLX.
The Agent Swarm feature scales to 300 simultaneous sub-agents across up to 4,000 coordinated steps. Moonshot tested it internally on long-horizon production tasks: the company used K2.6-based Claw Groups to run its own launch campaigns, with specialised agents for demos, benchmarks, social content, and video generation working in parallel.
Why It Matters
K2.6 is the first open-weight model to lead frontier proprietary models on HLE-Full with tools — a benchmark specifically designed to be resistant to contamination and to test real-world autonomous capability, not static knowledge. The SWE-Bench Pro lead over GPT-5.4 is 0.9 points; the HLE-Full margin over Claude Opus 4.6 is 1.0 point. Neither is a landslide, but both are in the right direction for a model that anyone can download and run.
Moonshot has held the open-weight top position against Chinese labs including DeepSeek and Alibaba throughout 2026. K2.6 is available now on Kimi.com, the Kimi App, the API, and the Kimi Code CLI.