Meituan’s LongCat team introduced LongCat-2.5 as a native multimodal model for long-horizon agent work. The headline numbers are large: about 1.6 trillion total parameters, roughly 48 billion active per token, and training that Meituan says ran entirely on domestic accelerators. Chat and API access are live; open weights were described as coming next.

Coffee summary

  • FACT: Meituan’s LongCat blog (dated 2026-11-1) and Meituan news (2026-11-02) describe LongCat-2.5 as a 1.6T-parameter MoE multimodal model with about 48B parameters activated per token.
  • FACT: The company says mid-training used sequences up to 1M tokens, and that training and W4A4 inference were optimized for domestic AI accelerators.
  • CLAIM: Meituan reports benchmark scores that sit between Claude Opus 4.8 and Claude Opus 5 on several of its tables; those figures are company-reported, not third-party audited.
  • FACT: Users can try the model on longcat.ai, and developers can call it through Meituan’s API; GitHub and Hugging Face releases were described as planned.
  • OPINION: Treat the release as a serious China-stack coding agent story, but verify any benchmark that matters to your buying decision.

What happened

LongCat-2.5 builds on LongCat-2.0 with two emphasis areas: native vision and long-horizon agentic coding. Vision inputs become multi-level discrete tokens so the same backbone can handle text and images. For agents, Meituan describes large-scale reinforcement learning on long trajectories, including turning real agent failures into verifiable training tasks.

Infrastructure claims are unusually specific. Meituan says domestic cards with under 80 GB of memory forced supernode-aware parallelism and dual-shard FSDP, and that a W4A4 scheme (4-bit weights and activations) improved prefill and decode throughput versus a W8A8 baseline in its own tests.

Why it matters

Most English coverage of frontier coding agents still centers on US and European labs. LongCat-2.5 is Meituan’s bid to show a trillion-scale multimodal coding model trained end-to-end on Chinese silicon, with an API already open. That matters for teams watching “sovereign” compute stacks, and for developers comparing open and closed coding agents.

It also matters because the marketing puts LongCat next to Claude Opus-class models. Without independent harnesses, those comparisons are a starting point for research, not a purchase verdict.

What changed, at a glance

Spec LongCat-2.5 (company)
Total / active params ~1.6T / ~48B per token
Focus Long-horizon agents, visual coding, 3D/design demos
Context training Up to 1M-token sequences (company)
Hardware story Trained on domestic accelerators; W4A4 inference
Access longcat.ai chat + API; open weights planned

Who should care

  • Engineers evaluating coding agents who want another API option beyond the usual US labs.
  • Researchers tracking China-trained MoE models and domestic-accelerator training recipes.
  • Product teams that need visual coding (UI from screenshots or design mocks) rather than text-only codegen.

Limitations and open questions

  • Benchmarks are Meituan’s. The blog table mixes self-run and starred public scores for other models. Re-run anything you care about.
  • Open weights are not out yet. Availability today is chat and API; Hugging Face/GitHub timing was not a firm date in the announcement.
  • Harness sensitivity. Coding agent scores swing with tools, max context, and judge models. Cross-lab tables are hard to read at face value.
  • Pricing can change. LongCat’s pay-as-you-go page listed promotional rates; confirm live platform prices before budgeting.

What to do next

  1. Try one real task on longcat.ai or the API: a multi-file refactor or a visual UI rebuild, not a toy hello-world.
  2. Label every score CLAIM until you reproduce it in your own harness.
  3. Watch for weight release if you need offline or fine-tune access.
  4. Compare against your current agent on the same repo and the same time budget.
  5. Read the failure story. Meituan’s own pipeline converts agent failures into training tasks—useful signal about where the model still struggles.

AIImpish Take

LongCat-2.5 is less interesting as a parameter count and more interesting as a full-stack claim: multimodal tokens, long-horizon RL, and domestic-chip training with an API already shipping. Use it, measure it, and keep Meituan’s leaderboard lines clearly marked as company claims.

Related on AIImpish

Sources: Introducing LongCat-2.5; Meituan news on LongCat-2.5; LongCat-2.5 API pricing.