Meituan open-sources LongCat-2.0, a 1.6T-parameter coding model trained entirely on Chinese chips

Chinese delivery-app giant Meituan open-sourced LongCat-2.0 on June 30 — a 1.6-trillion-parameter coding model it says is the first of its scale trained and served end-to-end on domestic Chinese hardware, with no Nvidia GPUs involved anywhere in the pipeline.

What's new

Per VentureBeat's report on the release, LongCat-2.0 is "the 1.6-trillion-parameter Mixture-of-Experts (MoE) system," activating only a fraction of that at inference time — "limiting active computation to an average of 48 billion parameters per token" — alongside "a functional 1-million-token context window." Meituan trained it "entirely on a cluster of over 50,000 domestic Chinese Application-Specific Integrated Circuits (ASICs)," and released the weights under a "commercially viable MIT license."

On benchmarks, Meituan reports LongCat-2.0 "registers an empirical 59.5 on SWE-bench Pro, surpassing GPT-5.5's benchmark of 58.6," plus "70.8 on Terminal-Bench 2.1, a 77.3 on SWE-bench Multilingual, and a 73.2 on the general corporate workflow simulator FORTE" — placing it ahead of GPT-5.5 and Gemini 3.1 Pro on that particular coding benchmark, though still behind Claude Opus 4.7 and 4.8.

Perhaps the most striking detail: LongCat-2.0 wasn't a surprise debut. It's the model that had been quietly running under the stealth codename "Owl Alpha" on OpenRouter for two months. During that stretch, "Owl Alpha accounted for approximately 10.1 trillion monthly tokens — averaging 559 billion tokens per day — representing a 242% month-over-month explosion in volume that propelled it into the platform's global top three," securing "the top ranking on the Hermes Agent workspace, second place on Claude Code deployments, and third place across international OpenClaw environments" — all before anyone knew who built it.

Context

LongCat-2.0 lands alongside a wave of Chinese open-weight coding models (DeepSeek, Qwen, Kimi/Moonshot, Z.ai/GLM) competing directly with GPT, Claude, and Gemini on agentic coding tasks, typically at a fraction of the API price — Meituan is pricing standard access at $0.75 per million input tokens and $2.95 per million output, discounted to $0.30/$1.20 during launch. What sets this release apart is the hardware story: Chinese AI labs have trained large models on non-Nvidia silicon before, but Meituan's claim of a full trillion-parameter-scale training-and-inference pipeline on a 50,000-card domestic ASIC cluster — with no Nvidia or Google TPU hardware anywhere in the process — is a specific, larger step toward supply-chain independence than prior domestic-chip efforts.

Why it matters

If Meituan's hardware claims hold up to scrutiny, LongCat-2.0 is evidence that US export controls on advanced chips are not fully containing the frontier-adjacent end of Chinese AI development — a near-frontier coding model was trained and served at scale without the Nvidia GPUs those controls were designed to restrict. The stealth-launch angle matters too: two months of unattributed top-three OpenRouter traffic under "Owl Alpha" means developers were already routing real production workloads to this model before its origin, license, or training hardware were public, which will likely intensify scrutiny of how OpenRouter and similar aggregators disclose model provenance going forward.