Velocity
This Week's Stories
Meituan

Meituan

AI · Large Language Models · AI Chips

Launch

Meituan open-sources 1.6T-parameter LongCat-2.5 agent model, no Nvidia GPUs

November 9, 2026

Benchmarks now land within striking distance of top closed US models, suggesting China's chip ecosystem can support frontier-scale training end to end.

  • Meituan's LongCat team released the full weights of LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts model with about 48 billion active parameters per token, under the MIT license, trained entirely on domestic Chinese AI chips.
  • Meituan, the Chinese food-delivery and e-commerce giant, built LongCat-2.5 for AI agents that work through long, multi-step coding and tool-use tasks autonomously.
  • A self-hosting guide shows the quantized model deployed on a 194-chip Huawei Ascend Atlas A2 cluster (2 chips for images, 64 for prompt ingestion, 128 for response generation), with no Nvidia GPU instructions provided.
  • On Meituan's own benchmarks, LongCat-2.5 scored 88.4 on Terminal Bench v2.1, just behind Claude Opus 5's 89.1 and GPT-5.6 Sol's 88.8, and jumped to 64.6 on DeepSWE from 9.7 for its predecessor LongCat-2.0.
  • 4-bit (W4A4) quantization of both expert weights and activations lifted prefill throughput about 43% versus an 8-bit baseline, while a custom FSDP setup raised long-sequence fine-tuning throughput 86% per card versus standard data parallelism.
  • The model follows LongCat-2.0, which Meituan trained on a cluster of more than 50,000 domestic AI chips, extending a push by Chinese labs to build frontier models without Nvidia hardware.
  • By validating a full training-to-deployment pipeline on domestic chips at near-frontier performance, LongCat-2.5 signals that China's AI chip ecosystem may now support large-scale model development independent of Nvidia's supply chain.

Get the app

Stay Ahead With Velocity

Deep company profiles, investor context, and every original source behind this story — plus the next one, the moment it breaks.

Download on the App StoreGet it on Google Play

More This Week

View All →