Hands-on with Meituan LongCat-2.5: trained on domestic chips, near Opus 4.8 level
葬AI · wechat · 2026-11-11
A detailed hands-on review of Meituan's LongCat-2.5, run through Claude Code as the harness. The model was optimized for domestic GPUs with under 80GB VRAM per card, and its tech report honestly benchmarks against GPT-5.5 and GPT-5.6sol.
Overall verdict: roughly in the Opus 4.8 range, covering 70% of ordinary office work.
- Frontend: complex frontend tasks (animated delivery e-bike, interactive planet, Japanese miniature shrine diorama) are essentially solved given precise prompts.
- Long-horizon coding: built a full web music player (dark/light themes, vinyl spin, waveform art, synced lyrics, hand-drawn curve-to-audio synthesis) in one shot; reverse-engineered a game binary into TypeScript over 5 hours with good early-level fidelity but gaps in complex late scenes. Sustains 5-10 hour runs, though can fall into thinking loops late in tasks.
- Cowork: Excel DCF modeling, annual-report retrieval and PPT generation handle 80% of typical office worker needs with minor flaws.
The author closes that LLMs are a brutal manufacturing industry driven by compute and data, praising Meituan for skipping leaderboard hype, genuinely training on domestic chips, and giving away tokens.
More from coding & agent
- Gemini /teamwork agents caught live: four parallel task logs all running that second — RyanSpunk · 2026-11-13
- Codex CLI 0.168.0 adds model provider picker, side-by-side TUI workspace and animated GIF generation — github-actions[bot] · 2026-11-13
- Free MCP server lets Claude read Altium, Allegro, KiCad and Gerber PCB files directly — woodisruff · 2026-11-13
- a16z Partner Clones Open-Source Game with Fable 5.2 Agent, Which Finds Its Own Horse Animation Asset — FinanceYF5 · 2026-11-13
- OpenAI team shares how it runs and governs fleets of persistent agents internally — romainhuet · 2026-11-13
- Perplexity's Decider scanned 1B tokens across 7,000 agent sessions, found 14 security fixes — denisyarats · 2026-11-13