Kimi K3.1 Pro weights drop tomorrow: 2.8T params, 1M-token context
togethercompute · x · 2026-11-04
Together AI announced Kimi K3.1 Pro has arrived, with weights dropping tomorrow. On Nov 4, Moonshot AI's Feihu Tang joins Together AI for a deep dive into its architecture and training decisions.
K3.1 Pro is a post-trained upgrade to Kimi K3: a MoE model with 2.8T total / 104B active parameters and a 1M-token context window. Moonshot reports Terminal-Bench 4.0 rising from 17.2% on base K3 to 63.6%, with substantial gains in coding, 3D creation, financial research and other long-horizon agentic capabilities. The session also covers where base K3 hit limits in long-running agent workflows and lessons from building it.
Related event: Kimi K3.1 Pro Released with 2.8T Parameters, Weights to Be Open-Sourced(2 posts)→
More from coding & agent
- MCP Visual Language uses GLM-4.5V to extract and analyze code from screenshots — modelcontextprotocol · 2026-11-04
- AI Design Blueprint packages 10 principles and 16 laws for coding agents over MCP — modelcontextprotocol · 2026-11-04
- AI agents can beat the Pangram AI detector — but the arms race is tightening — soumitrashukla9 · 2026-11-04
- Free MCP registry lets agents find data sources and log what they'd pay for missing ones — agentdemandregistry · 2026-11-04
- Open-source macOS notch assistant uses GPT-6.1 API to render real-time interactive 3D explanations — OpenAIDevs · 2026-11-04
- Fennick is a local-first PKM app where Claude edits your notebook as a peer — thesaraharminta · 2026-11-04