Meituan releases LongCat-2.5: 1.6T-parameter native multimodal model beating Claude on agent benchmarks
智东西 · wechat · 2026-11-02
Meituan's LongCat team released LongCat-2.5, a 1.6T-parameter MoE native multimodal model (48B active per token) targeting long-horizon agent tasks. It leads Claude-Opus-5 and GPT-5.6-Sol on AutomationBench and LongCatClawBench, scores 38.5 on TerminalBench v4.0 (beating GPT-5.6-Sol, trailing Claude-Opus-5). Vision tokens and text share one backbone via RVQ discrete tokenization, enabling visual coding and 3D generation. Training uses Agentic RL over auto-generated GitHub PR environments with partial rollout and anti-reward-hacking rewards, and a supernode-retuned FSDP/EP/CP scheme lifts per-card throughput 86% over DDP on domestic accelerators with under 80GB VRAM.
Related event: Meituan Releases LongCat-2.5, a 1.6T-Parameter MoE Agentic Model(4 posts)→
More from Models
- Gemini 4 Argon ships, putting Google back in the frontier model race after a year-long lull — danshipper · 2026-11-03
- Hands-on audit punctures Astra's '0.016% weight programming' claim: it just disabled Qwen's thinking mode — teortaxesTex · 2026-11-03
- Leak: Unreleased astra6.2 builds a 3DS-style Mario Kart in 20 min under 1% quota — lxfater · 2026-11-03
- "Sonnet 5.5 used to be smarter": users sense capability regression — nlarusstone · 2026-11-03
- Apple Intelligence lands in China in December, powered by Alibaba's Qwen instead of ChatGPT — winer666 · 2026-11-03
- Only recent models seem to know "weight surgery" — are labs building Bayesian weight priors? — Tkaraletsos · 2026-11-03