Trending on X: Moonshot's Kimi K2.6 Lands Open-Weight, Trillion-Parameter, and Within a Hair of Claude Opus on Coding
What Happened
Moonshot released Kimi K2.6 under a modified MIT license — a 1T-parameter MoE (32B active) with 256K context and native multimodal input. Reported scores include 80.2% on SWE-Bench Verified, 54.0% on HLE-Full with tools, and 83.2% on BrowseComp, broadly matching Claude Opus 4.6 and GPT-5.4. The model supports 300-agent swarm orchestration and INT4 self-hosting.
My Take
This is what the open-weight thesis looked like on paper, only happening faster than almost anyone underwrote. The frontier labs spent two years arguing closed weights were the moat; in 2026, the moat is product, distribution, and reliability — not the weights themselves. For any business still paying Opus prices for code review, summarization, or internal tooling, the cost calculus changed this morning. Watch for at least one US foundation lab to "respond" with a price cut by Friday.
Read Original Source