SHELFCRITTER AI Digest

COMPANY

Meituan

companytopic-note

Overview

Meituan is a Chinese super-app operator (food delivery, local services, retail) whose in-house AI research arm shipped the LongCat-2.0 model on July 1, 2026 — a 1.6T-total / 33–56B active MoE trained on a 35T-token corpus, and the first frontier-scale pre-training run completed end-to-end on domestic Chinese ASICs (a 50,000-card Huawei Atlas-950 SuperPod cluster) without a single NVIDIA GPU on the primary path. Prior to LongCat-2.0, Meituan surfaced as a smaller open-weights entry (LongCat-1.x) inside the Chinese MoE cluster; the 2.0 release is the entry point at which Meituan crosses into the frontier-scale-training-substrate story.

Timeline

  • 2026-07-01-AI-Digest — Meituan open-sources LongCat-2.0 — 1.6T total / 33–56B active MoE, 35T-token training run, trained end-to-end on a 50,000-card Huawei Atlas-950 SuperPod cluster without a single NVIDIA GPU on the primary path. Benchmark placement: SWE-bench Pro 59.5 (ahead of Gemini 3.1 Pro and GPT-5.5) and Multilingual 77.3, still behind Claude Opus 4.7 / Claude Opus 4.8 on general-purpose scores. Meituan has not publicly named the ASIC vendor beyond the Atlas-950 platform reference — Huawei Ascend 910C is the community-attributed underlying silicon, but the company itself has declined to confirm. The framing worth softening from mainstream coverage: this is the first confirmed end-to-end frontier-scale training on domestic ASICs — prior Chinese-hardware announcements (DeepSeek V4-Pro, April 2026) were Huawei-post-trained on Nvidia-pre-trained lineage, and the failed mid-2025 full-Ascend attempts predate this cleanly. The “China can’t train frontier models without Nvidia” premise no longer survives contact with a public 1.6T open-weights release; the harder open question is whether the training-run economics (unnamed hardware cost, undisclosed cluster utilisation) close the gap on cost-per-token, not just on capability.

Key Developments

  1. First Frontier-Scale End-to-End Training on Domestic Chinese ASICs (July 1, 2026): LongCat-2.0 (1.6T MoE, 35T tokens) trained end-to-end on 50,000 Huawei Atlas-950 domestic ASICs is the first frontier-scale pre-training run to complete without a single NVIDIA GPU on the primary path. The precision points the corpus carries: confirmed end-to-end is the load-bearing framing (prior Chinese-hardware announcements were post-trained on Nvidia-pre-trained lineage), and Huawei Ascend 910C is the community-attributed silicon behind the Atlas-950 platform reference — Meituan has not confirmed the underlying chip. Capability demonstrated, not parity.

  2. Benchmark Placement — Ahead on Coding, Behind on General: SWE-bench Pro 59.5 (ahead of Gemini 3.1 Pro and GPT-5.5) and Multilingual 77.3, still behind Claude Opus 4.7 / Claude Opus 4.8 on general-purpose scores. The read that survives verification is that LongCat-2.0 leads on coding within its sparsity envelope while trailing the closed frontier on breadth, not that it has crossed the capability frontier as a whole.

  3. Training-Economics Question Is the Open Test, Not the Capability Question: The 60-day watch item is whether public evidence surfaces on cost-per-token and cluster utilisation of the 50,000-card Atlas-950 run. The capability question — can Chinese-domestic silicon train a 1.6T MoE end-to-end — has a public answer; the economics question — can it do so at competitive per-token cost — is the harder open question the corpus is now carrying.

See also: LongCat-2.0, Huawei, NVIDIA, DeepSeek, MOC - Open Source Models, MOC - AI Infrastructure.