Meituan Open-Sources 1.6T Agent Model Trained Without Nvidia GPUs

1 reports archived1 independent sourcesupdated 11/15/2026, 18:06:06

Synthesis & Latest Updates

1 Sources Cross-Validated

Meituan's LongCat team has open-sourced LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts model purpose-built for long-horizon autonomous agent tasks, especially coding, activating roughly 48 billion parameters per token with a one-million-token API context. Trained entirely on Chinese hardware without any Nvidia GPUs, it approaches the coding benchmark levels Western frontier models such as Claude Opus 5 reached a few months earlier.

LATEST/Meituan's LongCat team has open-sourced LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts model purpose-built for long-horizon autonomous agent tasks, especially coding, activating roughly 48 billion parameters per token with a one-million-token API context. Trained entirely on Chinese hardware without any Nvidia GPUs, it approaches the coding benchmark levels Western frontier models such as Claude Opus 5 reached a few months earlier.

Hourly heat curve

5 fully observed hours

Now

10.6

Peak

11.819:00

24h change

–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 1 reports · Latest first

The DecoderT2Nov 15, 18:06·88 pts
Meituan Open-Sources 1.6T Agent Model Trained Without Nvidia GPUs
- MoE architecture with 1.6T total and ~48B active parameters per token, plus a 1M-token API context
- Trained end-to-end without Nvidia GPUs, demonstrating domestic compute stacks can support frontier-scale training
- Meituan claims coding benchmarks near the level of Claude Opus 5 and peers from months earlier; weights are downloadable
View Related Reports & Analysis Original Source