Meituan launches LongCat-2.5: 1.6T-param native multimodal agentic model
thisdudelikesAI · x · 2026-11-04
Meituan released LongCat-2.5, a natively multimodal agentic model with 1.6T total parameters (48B active), built for long-horizon agent tasks with stronger visual understanding and agentic coding for complex projects, creative design, and 3D work. Key tech: visual primitive modeling (images as discrete tokens, vision as first-class capability), unified visual-text token space, and W4A4 inference cutting TTFT by 16%. Pricing: $0.3/M input (cache miss), $0.006/M (cache hit), $1.2/M output.
More from coding & agent
- Open-source barebox sandbox: r/w/x permissions let AI agents run with 1-second startup — Tight_Heron1730 · 2026-11-05
- GitHub Octoverse: 40%+ of PRs now started by agents, 17.9B pushes in a year — ggerganov · 2026-11-05
- Cherry Pi uses SQLite as substrate: apps, layers, and memory as single database files — sull · 2026-11-05
- 6 months of AI-assisted dev across 37 repos: the bottleneck is suspiciously human — Xander-XGCS · 2026-11-05
- Open-source MCP Master Tester: npx-run any MCP server with a tool-definition linter — simplymanas · 2026-11-05
- Infographic: the 7 layers of the LLM stack, from data acquisition to apps — nevrekaraishwa2 · 2026-11-05