OPEN LINK ↗
// 1h agoMODEL RELEASE
LongCat-2.5 launches multimodal coding agents
Meituan’s LongCat-2.5 is a 1.6-trillion-parameter MoE model with native vision, approximately 48 billion active parameters per token, long-context reasoning, and agentic coding. It targets full-project engineering, visual programming, creative design, and 3D workflows.
// ANALYSIS
LongCat-2.5’s key advance is treating multimodal input and extended agent execution as core model capabilities, not add-ons.
- –Native visual tokenization enables image-aware coding, design, and 3D tasks.
- –Long-horizon reinforcement learning focuses on completing multi-step engineering work across complex codebases.
- –Compatibility with Claude Code, Codex, Hermes, and other harnesses lowers adoption friction for developers.
- –Vendor-reported Terminal Bench 2.1 results improve substantially over LongCat-2.0, though independent validation is still needed.
- –W4A4 inference optimizations could make a trillion-parameter model more practical, with hardware-dependent throughput gains.
// TAGS
longcat-2.5llmmoemultimodallong-contextcoding-agentagentai-coding
DISCOVERED
1h ago
2026-11-03
PUBLISHED
1h ago
2026-11-03
RELEVANCE
9/ 10
AUTHOR
AI Revolution