Meituan LongCat Owl Alpha Tops OpenRouter — Trained Entirely on Domestic Chinese ASICs
> Published: 2026-06-29 / 06-30 · Category: Model Release/Update · Tags: AI sovereignty / model distribution > Sources: X: Emad Mostaque + official Meituan LongCat account > Original link: https://x.com/EMostaque/status/2071701921241448574 > Model blog: https://longcat.chat/blog/longcat-2.0/
What Happened
On June 29 at 9 PM (5 AM Beijing time on June 30), Emad Mostaque posted:
> "The most popular model on OpenRouter (10 trillion tokens) is a 1.6 trillion parameter MoE from Meituan (China's super-app / DoorDash). Basically Gemini / Opus 4.6 level. 35 trillion tokens of training, all on 50k domestic ASICs. No GPUs needed."
Meituan's official LongCat account confirmed nine hours later:
> "Owl Alpha on OpenRouter — that's ours. Since public beta launch, DAU has reached global Top 3, #1 on Hermes Agent, #2 on Claude Code, #3 monthly on OpenClaw."
Public Information Summary
- Architecture: 1.6T total-parameter MoE (activation count undisclosed, presumably in the tens of billions);
- Training data: 35 trillion tokens;
- Training compute: deployed entirely on 50,000 domestic ASICs (model undisclosed; timing suggests Huawei Ascend, Cambricon, or a next-gen flagship from another domestic ASIC vendor);
- Performance positioning: Emad called it "Gemini / Opus 4.6 level" — likely an exaggeration — but the 10 trillion tokens of OpenRouter usage is a hard metric of repeated production use;
- Distribution: OpenRouter (neutral international platform);
- Use cases: per official claims, #1 / #2 / #3 on Hermes Agent (open-source coding agent), Claude Code (Anthropic's coding agent), and OpenClaw (multimodal agent) respectively.
- In the GPU era, this training would need roughly 1,000+ H100/H200s running weeks to months;
- Doing it on domestic ASICs means the domestic AI compute stack has completed software-level adaptation for "10K-card-class MoE training" — distributed training frameworks, all-reduce communication libraries, FP8/BF16 mixed precision, checkpoint fault tolerance all work;
- The "50,000 ASICs" figure is interesting — if Huawei Ascend, per-card compute is weaker than H100 but compensated by scale; if a more aggressive next-gen ASIC (Moore Threads, MetaX, Cambricon), it may mean per-card compute matching H100.
- Western developers switching models on OpenRouter will genuinely "try" this Chinese model;
- In coding agent workflows, if the model is "good enough," developers have no incentive to switch back — trials become DAU;
- This is a channel entirely different from domestic distribution — it bypasses traditional regulatory friction in Chinese models going overseas.
- 06-26 · OpenRouter MCP server launch — OpenRouter building an agent-era LLM gateway;
- 06-28 · Wayfinder Router — deterministic routing with no model calls;
- 06-29 · Claude apps gateway on Bedrock and Google Cloud — Anthropic expanding its own distribution;
- 06-30 · LongCat Owl Alpha #1 on OpenRouter — real penetration of a domestic model on a Western neutral channel.
- Hermes Agent #1 — open-source coding agent; LongCat is strongest in the open-source agent ecosystem;
- Claude Code #2 — Anthropic's own coding agent. A notable signal: Anthropic users actively switched the backend model to LongCat Owl Alpha, implying coding performance at least comparable to Claude Opus 4.6, or users wouldn't switch;
- OpenClaw #3 — multimodal agent, indicating multimodal usability.
- The 50,000-ASIC vendor is undisclosed — Huawei Ascend, Cambricon, or a newer domestic flagship? This determines how real the "compute autonomy" claim is;
- Training data composition (35T tokens) undisclosed — English data? Synthetic data? Quality directly affects real English-language performance;
- Activation parameters of the 1.6T MoE undisclosed — if activation is very small (e.g., under 20B), "1.6T" is a marketing number with inference speed no better than a 27B dense model;
- Whether OpenRouter's 10T tokens is cumulative or recent is unclear — cumulative indicates long-term reputation; recent figures indicate short-term hype;
- Emad's "Gemini / Opus 4.6 level" needs discounting — the real level may sit between Sonnet 4.5 and Opus 4.5;
- Whether Anthropic will respond to users switching Claude Code's backend to LongCat — a signal worth watching for ecosystem dynamics.
This is the first reported case of a trillion-scale model trained entirely on domestic Chinese ASICs reaching top-tier usage on a neutral international distribution channel.
Deep Dive
Why is this heavier than it looks?
Over the past two years, Chinese LLMs going global followed the path of "Chinese alternatives to GPT / Claude" — GLM, Qwen, DeepSeek, Kimi could match GPT-4 on international benchmarks, but their real usage on Western mainstream distribution channels remained a minority option.
LongCat Owl Alpha is different. Its usage scenario is coding agents + coding CLIs — a toolchain Western developers use daily (Hermes, Claude Code, OpenClaw). Holding a Top 3 spot inside Western developer workflows is a genuine validation of the domestic model ecosystem.
"Gemini / Opus 4.6 level" should be discounted — Emad's taste runs aggressive. But 10 trillion tokens of real OpenRouter usage isn't discounted: it means thousands of Western developers use this model daily to write code.
What does "trillion-scale model on domestic ASICs" mean?
Technical detail: 50,000 ASICs training a 1.6T MoE on 35T tokens. No GPU involvement.
Either way, this validates that the effectiveness of compute restrictions in large-model training is rapidly decaying.
OpenRouter's neutral platform role.
OpenRouter, launched in 2024, is an LLM API aggregator — it connects OpenAI, Anthropic, Google, Mistral, Meta, and various Chinese models behind a unified OpenAI-compatible API. LongCat Owl Alpha reaching #1 means:
Relation to earlier signals:
Together these point one direction: the "LLM gateway" is becoming the new "app store" — whoever owns this distribution layer controls AI-era channels. LongCat Owl Alpha is the first domestic model to reach the top of a Western neutral distribution channel.
Use-case details.
Why It Matters
1. "Trillion-scale model on domestic ASICs" is a key 2026 milestone for AI infrastructure self-sufficiency — it validates the decay of compute-restriction effectiveness, far more significant than another model release. 2. "Chinese models at the top of Western neutral distribution" is a new going-global paradigm — neutral platforms like OpenRouter offer a more direct channel than the old "overseas subsidiary + overseas partner" route. 3. Model replaceability in the coding-agent era is real — switching Claude Code's backend from Opus to LongCat shows model choice is driven by actual results, not brand loyalty — a stress test for every model company. 4. Meituan's AI strategy has been seriously underestimated — a 1.6T MoE trained on domestic ASICs means Meituan's AI capability is no longer bound to any external vendor. 5. OpenRouter's channel value may matter more than the model itself — whoever owns AI's "app store" controls the next decade of distribution.