Moonshot releases Kimi K3, a 2.8T open-weight multimodal agent with 1M context
johnseach · x · 2026-07-28
- Moonshot AI has released Kimi K3 on Hugging Face as an open-weight, native multimodal agentic model.
- The model is described as Moonshot’s most capable system yet: 2.8T parameters, built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes).
- It supports a 1-million-token context window, native vision capabilities, and is positioned for long-horizon coding, knowledge work, and reasoning.
- Moonshot says K3 uses a Stable LatentMoE setup that activates 16 of 896 experts, which it claims improves scaling efficiency by about 2.5× versus Kimi K2.
- The release emphasizes agentic workflows such as terminal tool use, repository-scale coding, deep research, dashboards, motion design, video editing, and even chip design use cases.
Related event: Moonshot AI Releases Open-Weight Kimi K3 Model(41 posts)→
More from coding & agent
- Squarespace shows how AI can 'stain the page' for faster prototyping — rseroter · 2026-07-28
- Bittensor gets an MCP-style front door for agents to route storage, inference and training — markjeffrey · 2026-07-28
- Codex session audit asks for recurring long-run failures and workflow fixes — kevinkern · 2026-07-28
- Codex session audit shows sandbox confusion can cascade into worse mistakes — kevinkern · 2026-07-28
- Kimi K3 paper describes a three-stage post-training stack with SFT, RL, and MOPD — stochasticchasm · 2026-07-28
- Coding Agents Exhibit Weird Emergent Behavior: Leaving Escape Notes for Future Selves — chris_j_paxton · 2026-07-28