Moonshot Releases KIMI K3: The World's Largest Open-Source Model
Moonshot unveils KIMI K3, a 2.8 trillion parameter MoE model with a 1M token context window, outperforming major models in knowledge-based tasks.
Moonshot has just released the largest open-source model in history — KIMI K3.
> 2.8 trillion parameters
>MoE architecture with 896 experts, 16 of which are active
>1 million token context window
>Native image understanding
>New Kimi Delta Attention + Attention Residuals architecture
>Approximately 2.5 times higher scaling efficiency compared to K2
>Third most intelligent model after Fable 5 and GPT-5.6
>Outperforms Opus 4.8 MAX in real-world intellectual and knowledge-based tasks, according to the GDPval benchmark
>New SOTA result — 91.2 points in BrowseComp without context tricks
Kimi claims that K3 competes with Opus and sets a corresponding price: $3.00 per 1 million input tokens. $15.00 per 1 million output tokens.
Full open weights of the model will be published in a few days.
Comments
Weekly digest
The best of vibe coding, AI agents and open source — once a week, no spam.
Related articles
All articlesGLM-5.3 Flash Released: Zhipu's New AI Model Beats GLM-5.2
Zhipu officially releases GLM-5.3 Flash, scoring 84.3 on Terminal-Bench, outperforming GLM-5.2 at 10x lower cost. Open weights, MIT license, 1M context.
Cline Desktop Launches AI Agent Workspace for macOS & Windows
Cline Desktop offers an open‑source AI workspace with parallel agents, task scheduling, checkpoints, branching, and a plugin marketplace, now on macOS and Windows.
MiniMax Code Goes Open Source: Performance Beats Codex & Claude
MiniMax Code is now open on GitHub, offering faster, cheaper AI coding with 76.7% success—outperforming Codex and Claude Code.