Moonshot Releases Kimi K3: 2.8T Parameter MoE Model on Hugging Face
Moonshot releases Kimi K3 weights: a 2.8T parameter MoE model with 1M token context window, image understanding, and advanced deployment infrastructure.
Moonshot, as promised, has released the Kimi K3 weights
The model is already on Hugging Face.
Kimi K3 is their most powerful model: 2.8 trillion parameters in an MoE architecture, native image understanding, and a context window of 1 million tokens.
Moreover, they didn't just release the weights. Moonshot also released part of the K3 stack: high-performance attention kernels, a communication library for MoE, and infrastructure for large-scale deployment of agentic environments.
Comments
Weekly digest
The best of vibe coding, AI agents and open source — once a week, no spam.
Related articles
All articlesDeepSeek V4-Pro: Open Models, OpenAI API, Lower Off-Peak Pricing
DeepSeek V4-Pro matches top open models, nears GPT-5.6. New features: OpenAI Responses API, three reasoning levels. Pricing: peak/off-peak with 50% off-peak discount. Available Aug 16, 2026.
DeepSeek V4.1 Flash 1017: 244 tps, $500/mo, Free API Search Till Oct 31
DeepSeek V4.1 Flash 1017 delivers 244 tokens per second, beating GPT‑6.1 Sol, priced at $500/month. Its web‑search API is free until Oct 31, then $0.75 per 1k requests.
OpenAI Tests New Model by Rewriting Codex in Rust for Cross‑Platform
OpenAI’s Andrew Ambrosino says each new model must rewrite the desktop Codex in Rust and GPUI, aiming for a single model that can build native Mac, Windows and cross‑platform apps.