Moonshot AI open-sourced Kimi K2.6 weights on Hugging Face on April 20. It scored 58.6 on SWE-Bench Pro — ahead of GPT-5.4 (xhigh) at 57.7, Claude Opus 4.6 (max effort) at 53.4, and Gemini 3.1 Pro (thinking high) at 54.2. First open-source coder to clear the closed flagships on that bench.
What the model actually is
Kimi K2.6 is a 1T-parameter MoE with 32B active and a 262k context window. Beyond the headline number, it posts SWE-Bench Verified 80.2, BrowseComp 83.2, and Terminal-Bench 2.0 at 66.7. The real flex is agent swarming — coordinating 300 parallel sub-agents across 4,000 steps. In one published run it autonomously refactored an 8-year-old open-source financial matching engine over 13 hours, lifting median throughput 185% and peak 133%.
How to plug it in
Weights are on Hugging Face under an open license. Cloudflare Workers AI and DeepInfra are already serving it; Moonshot also runs a hosted API. So you can self-host the 32B active footprint, hit a managed endpoint, or wire it into agent frameworks that need long-horizon coordination — refactor pipelines, automated research loops, multi-step trading bots. HN bumped it back to the front page on April 30 after a “Word Gem Puzzle” coding contest where K2.6 outscored Claude, GPT-5.5, and Gemini.
You Might Also Like
- Mistral Medium 3 5 Scores 77 6 on swe Bench one Point shy of Gemini 3 1 pro
- Hiveterm Bets on the Multi Agent Workspace Claude Codex and Gemini in one Terminal
- Gpt oss 120b Openai Finally Goes Open Source and its Worth the Wait
- Claude Code Remote Control Just Turned my Phone Into a Coding Terminal and im Weirdly Into it
- Hermes Agent by Nous Research Might be the Open Source ai Agent That Finally Remembers Everything
Top AI Product 
Leave a comment