Meituan Open-Sources LongCat-2.0, a 1.6T MoE Trained Entirely Without NVIDIA Hardware
Weights and inference code shipped under MIT on July 5 for what Meituan calls the first trillion-parameter model trained and served end-to-end on non-NVIDIA silicon — with export-control implications in both directions.
Meituan released the weights and inference code for LongCat-2.0 under an MIT license on Hugging Face on July 5, 2026, following a June 30 announcement: a 1.6-trillion-total-parameter MoE with roughly 48B active parameters and native 1M-token context, trained on more than 50,000 domestic AI ASICs over 35 trillion tokens, per NYU Shanghai's RITS analysis.
Context
Frontier-scale training has been an overwhelmingly NVIDIA story. Meituan claims LongCat-2.0 is "the industry's first trillion-parameter model to complete full-process training and inference" on alternative hardware — with no rollbacks or irrecoverable loss spikes across the run. The silicon is suspected to be Huawei Ascend-class, though that remains unconfirmed.
What changed
- Open weights, MIT license: live at `meituan-longcat/LongCat-2.0` on Hugging Face since July 5, with inference code.
- Architecture: LongCat Sparse Attention enables native 1M-token context.
- Vendor benchmarks: 70.8 on Terminal-Bench 2.1 and 59.5 on SWE-bench Pro — marginally ahead of GPT-5.5's 58.6 on the latter, per Meituan's numbers.
- Stealth track record: the model had been serving anonymously as "Owl Alpha" atop OpenRouter usage charts before disclosure.
Why it matters
If the claim holds, frontier training is no longer exclusively an NVIDIA story — evidence that US export controls have catalyzed a viable domestic training stack in China. That cuts both ways for policy: it demonstrates Chinese self-sufficiency while also handing Western buyers an MIT-licensed trillion-parameter model. Deployment is datacenter-only, however: the reference configuration is two nodes of 16 H20-class accelerators each.
Details
The release was corroborated by explainx.ai coverage updated through July 5–6. "Owl Alpha"'s OpenRouter performance provides an indirect, usage-based signal of real-world competitiveness.
Deployment reality
The MIT license makes LongCat-2.0 genuinely usable by Western enterprises and inference providers — no attribution-clause triggers of the kind attached to Kimi K2.7's Modified MIT — but the hardware requirements gate who can actually run it. The practical near-term beneficiaries are clouds and inference providers with non-consumer clusters, echoing the pattern across July's open-weight wave: weights are free, compute is not. The "Owl Alpha" episode also raises provenance questions buyers should weigh: the model accumulated real production traffic before its identity was public.
Limitations and caveats
The non-NVIDIA training claim is Meituan's own; the specific silicon has not been confirmed, and benchmark figures are vendor-reported. The reference deployment's reliance on H20-class parts — themselves export-controlled NVIDIA products — complicates a clean "NVIDIA-free" framing.
Sources
*Update note: This post was last reviewed on 2026-07-22. Independent evaluation of the released weights and confirmation of the training hardware are open verification items.*
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.