Moonshot releases Kimi K3, a 2.8T open-weight multimodal agent with 1M context
victormustar · x · 2026-07-28
Moonshot AI has released Kimi K3 as an open-weight, native multimodal agentic model with 2.8T parameters and a 1-million-token context window.
The model card says Kimi K3 is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), using a Stable LatentMoE design that activates 16 of 896 experts and improves scaling efficiency by about 2.5× over Kimi K2. Moonshot positions it for long-horizon coding, knowledge work, and multimodal reasoning, and says the weights are fully released under the Kimi K3 License.
Related event: Moonshot Releases Open-Weight Kimi K3 Model(147 posts)→
More from coding & agent
- A remote MCP server adds OAuth 2.1 and verified auto-apply tools for job workflows — tom_doerr · 2026-07-28
- Clients pay for an AI slide maker that enforces template consistency, not creativity — Odd_Report6798 · 2026-07-28
- A single SKILL.md adds a local Python sandbox for data-analysis agents — dheetoo · 2026-07-28
- MCP in Practice: Building a Job Application Server with OAuth 2.1 — Shape_Weird · 2026-07-28
- VerifiMind PEAS ships a multi-agent validation stack with 13 free tools — modelcontextprotocol · 2026-07-28
- Directus MCP Server turns Directus into a full AI-accessible backend — modelcontextprotocol · 2026-07-28