SHELFCRITTER AI Digest

MODEL

LongCat-2.0

modeltopic-noteopen-source

Overview

LongCat-2.0 is Meituan‘s July 2026 frontier-scale open-weights MoE — 1.6T total parameters / 33–56B active — trained end-to-end on a 50,000-card Huawei Atlas-950 SuperPod cluster across a 35T-token corpus without a single NVIDIA GPU on the primary path. It is the first frontier-scale pre-training run completed on domestic Chinese ASICs, and the release marks the transition of the “China can’t train frontier models without Nvidia” premise from live debate to a public 1.6T open-weights counter-example.

Timeline

  • 2026-07-01-AI-Digest — Meituan open-sources LongCat-2.0 (1.6T total / 33–56B active MoE, 35T-token training) with full architectural and training details published. 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 narrow read is capability demonstrated, not parity; the structural read is that 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. Extends the 2026-06-30-AI-Digest high-sparsity MoE cluster note (LongCat surfaced there as an HN item) with the training-substrate detail that reframes it.

Key Developments

  1. First Frontier-Scale End-to-End Training on Domestic Chinese ASICs: 1.6T MoE trained end-to-end on 50,000 Huawei Atlas-950 domestic ASICs across a 35T-token corpus. 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.

  2. High-Sparsity Trillion-Total MoE as Convergent Architecture Pattern: LongCat-2.0’s 1.6T total / 33–56B active envelope joins DeepSeek V4 Pro and Kimi K2.5 as the third release inside a comparable window sharing the same sparsity shape. What was “one-off” in Q1 is now visibly convergent architecture choice across three independent Chinese labs.

  3. Coding-Ahead, General-Behind 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. 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.

  4. Training-Economics Question Is the Open Test: 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. Capability is now a public data point; economics remains the harder open question the corpus is carrying forward.

See also: Meituan, Huawei, DeepSeek V4 Pro, NVIDIA, Claude Opus 4.8, MOC - Open Source Models, MOC - AI Infrastructure.