LAUNCHES

Meituan’s LongCat-2.5 Agent Model Was Trained Without Any Nvidia GPUs

R Ryan Matsuda Nov 15, 2026 2 min read
Engine Score 8/10 — Important

1.6T-param agent model built without Nvidia GPUs is high-impact counter-narrative to export-control story

Editorial illustration for: Meituan's LongCat-2.5 Agent Model Was Trained Without Any Nvidia GPUs
  • Meituan’s LongCat team released open weights for LongCat-2.5, a 1.6-trillion-parameter agent model, one week after its official unveiling.
  • The mixture-of-experts model activates roughly 48 billion of its 1.6 trillion parameters per token and was trained entirely on Chinese hardware.
  • LongCat-2.5 handles up to 1 million tokens of context, generates responses up to 128,000 tokens long, and tokenizes images alongside text instead of using a bolted-on image encoder.
  • Meituan trained the model’s coding-agent skills on real GitHub pull requests inside environments that blocked internet access and screened for shortcuts.

What Happened

The LongCat team at Chinese food-delivery company Meituan released the open weights for LongCat-2.5 on November 15, 2026, one week after the model’s official unveiling, The Decoder reported the same day. Meituan says the 1.6-trillion-parameter model was trained entirely on Chinese hardware and built specifically for agents working through long, multi-step coding tasks.

Why It Matters

LongCat-2.5 arrives as Nvidia’s GPU-export restrictions remain a live flashpoint; a forensic analysis recently found Nvidia Blackwell fingerprints inside six Chinese AI models despite export controls, making Meituan’s claim of fully domestic-hardware training a notable counterpoint. Meituan says LongCat-2.5 comes close to the level Western models like Claude Opus 5 reached on coding benchmarks a few months ago.

Technical Details

LongCat-2.5 is a mixture-of-experts model that activates about 48 billion of its 1.6 trillion total parameters per token. Through the API it handles up to 1 million tokens of context and generates responses as long as 128,000 tokens. Unlike models that bolt on a separate image encoder, it converts images into tokens processed alongside text, an approach drawn from Meituan’s research model LongCat-Next. For reinforcement learning, Meituan’s team used real software tasks pulled from GitHub pull requests, blocking internet access and removing Git history in training environments, with a language model checking solutions for shortcuts.

Who’s Affected

Enterprises and developers evaluating open-weight coding agents gain another option that Meituan positions as competitive with top Western models without relying on Nvidia hardware, a pitch aimed squarely at Chinese AI labs navigating US chip export controls.

What’s Next

Meituan’s benchmark comparisons to Claude Opus 5 are self-reported; independent researchers have not yet published third-party verification of LongCat-2.5’s coding performance.

Related Reading

Share

Enjoyed this story?

Get articles like this delivered daily. The Engine Room — free AI intelligence newsletter.

Free · No spam · Unsubscribe anytime