China Hot Tech @ChinaHotTech
← All posts
Models · API · Agentic AI

Meituan LongCat-2.5: 1.6T MoE Agentic Multimodal API

LongCat-2.5 from 美团 Meituan is a native multimodal model aimed at long-horizon agentic work — coding across whole projects, interactive web / 3D / visual coding, and end-to-end engineering loops. The headline shape is a 1.6 trillion parameter MoE with roughly 48B activated per token, building on LongCat-2.0 with stronger native visual understanding. You call it as an API product (OpenAI- and Anthropic-compatible endpoints, model id LongCat-2.5, 1M context / 128K max output) — not as open weights, and not as a Digi-Key SKU. Vendor benches and accelerator throughput claims below are flagged with asterisks.

· ~7 min read · China Hot Tech

Soft saffron, warm cream, deep ink charcoal and coral accent editorial: agent coding terminal meets soft print — Meituan LongCat-2.5 multimodal API

Illustration: soft saffron / warm cream / deep ink charcoal + coral accent — agent coding terminal meets soft print editorial. Distinct from morning Iluvatar steel-blue/teal and from MiniCPM lilac/jade, Xiaomi paper/coral, UniWAM fog-blue/amber, DF1000 sand/copper, X-Era mint/cream, Kling cream/cyan. Asterisked labels are vendor / press figures.

Frontier agent stacks already wire Claude or GPT through OpenAI-shaped clients. LongCat-2.5’s overseas-relevant move is simpler than another China-only chat UI: Meituan published OpenAI-compatible and Anthropic-compatible HTTPS bases you can point existing SDKs at, with a 1M-token context window and a multimodal + agent-coding pitch. The model was trained and mid-trained for long context on domestic AI accelerators (“国芯国模”), and Meituan’s own tables put many agent / code / visual benches between Claude-Opus-4.8 and Claude-Opus-5*. Treat that as a vendor leaderboard signal, then decide whether an API key at longcat.ai belongs in your eval harness.

What shipped

Per Meituan’s LongCat-2.5 blog and platform docs (release window ~2026-11-01/02):

Can overseas teams use it today?

As an API — yes, if you can create an account and mint a key on the LongCat platform. Bases: OpenAI-compatible https://api.longcat.chat/openai and Anthropic-compatible https://api.longcat.chat/anthropic; model id LongCat-2.5. Try / console at longcat.ai; docs at longcat.ai/platform/docs/zh/. Meituan also claims harness compatibility with Claude Code, Codex, Hermes, and similar agent tools*. As open weights / a Hugging Face download — no. Pricing and geo availability can vary; verify signup and billing yourself before baking it into production.

Editorial summary panels: 1.6T MoE ~48B active, 1M context, OpenAI and Anthropic-compatible API bases, W4A4 domestic-accelerator throughput asterisks, agent coding + native vision — LongCat-2.5

API shape + agent/vision pitch + domestic-accelerator throughput* — the three overseas checks. Asterisks mark vendor / Meituan / press figures.

API surface

The developer path is deliberately SDK-familiar:

That means you can usually keep your existing OpenAI or Anthropic client libraries and swap base URL + model string (plus key). Official blog: longcat.ai/blog/longcat-2.5. Chinese press coverage includes 经济观察网 (EEO, 2026-11-02).

Benches* (vendor table)

Meituan’s published comparisons place LongCat-2.5 between Claude-Opus-4.8 and Claude-Opus-5 on many agent / code / visual benches*. Those are vendor tables, not independent re-runs. Use them as a shortlist signal for agent and multimodal coding evals — then measure on your own harness (repo tasks, visual web flows, long-context tool loops).

What changes for an overseas stack

Separately, Meituan is wiring the line into merchant ops via the “妙手” AI console for local-life brands (古茗 and peers). That is a China product surface — useful context, not the overseas integration hook.

Caveats

Bottom line: LongCat-2.5 is Meituan’s bid at a frontier agentic multimodal API you can hit with OpenAI or Anthropic clients — 1.6T MoE, ~48B active, 1M context, native vision, long-horizon coding RL — as long as you keep asterisks on every vendor bench and accelerator claim, and treat it as a metered API rather than a downloadable model.

Sources