> Published: 2026-06-29 / 06-30 · Category: Model release/update · Tags: AI sovereignty / Model distribution > Source: X: Emad Mostaque + the official Meituan_LongCat account > Original post: https://x.com/EMostaque/status/2071701921241448574 > Model blog: https://longcat.chat/blog/longcat-2.0/
What Happened
At 9 PM on June 29 (5 AM Beijing time, June 30), Emad Mostaque — one of AI's most-talked-about commentators — posted:
> "The most popular model on OpenRouter (10 trillion tokens) turns out to be a 1.6T parameter MoE from Meituan (China's super-app / DoorDash). Basically Gemini / Opus 4.6 level. Trained on 35 trillion tokens, entirely on 50,000 domestic ASICs. No GPUs needed."
Meituan's LongCat official account confirmed nine hours later (morning of June 30):
> "Owl Alpha on OpenRouter — that's ours. Since public beta launch, daily active usage has reached global Top 3, #1 on Hermes Agent, #2 on Claude Code, and monthly #3 on OpenClaw."
Summary of public information:
- Architecture: 1.6T total parameters MoE (activated parameters undisclosed; speculated to be in the tens of billions);
- Training data: 35 trillion tokens;
- Training compute: entirely on 50,000 domestic Chinese ASICs (model undisclosed; based on timing, speculated to be a next-gen flagship from Huawei Ascend, Cambricon, or another domestic ASIC vendor);
- Performance positioning: Emad called it "Gemini / Opus 4.6 level" — likely exaggerated, but 10 trillion tokens of real usage on OpenRouter is a hard metric, meaning it is being repeatedly used in real production scenarios;
- Distribution: OpenRouter (a neutral international platform);
- Use cases: per official claims, #1 on Hermes Agent (open-source coding agent), #2 on Claude Code (Anthropic's coding agent), #3 on OpenClaw (multimodal agent).
- In the GPU era, 1.6T MoE training compute would require roughly 1,000+ H100/H200s running for weeks to months;
- Doing the same on domestic ASICs means the domestic AI compute stack has completed software-level adaptation for '10,000-card-class MoE training' — distributed training frameworks, all-reduce communication libraries, FP8/BF16 mixed precision, checkpoint fault tolerance — the whole software stack works;
- The "50,000 domestic ASICs" figure itself is interesting. If they are Huawei Ascend, per-card compute is weaker than H100 but compensated by scale; if they are a more aggressive next-gen ASIC (Moore Threads, MetaX, Cambricon), it could mean "per-card compute on par with H100."
- Western developers switching models via OpenRouter will genuinely 'try' this Chinese model;
- In coding agent workflows, if the model is 'good enough,' developers have no incentive to switch back — 'trial' becomes 'daily active';
- This is a channel entirely different from domestic Chinese distribution — it bypasses the traditional regulatory friction of "Chinese LLMs going overseas."
- 06-26 · OpenRouter MCP server launch — OpenRouter itself building an agent-era LLM gateway;
- 06-28 · Wayfinder Router — deterministic routing without calling any model;
- 06-29 · Claude apps gateway on Bedrock and Google Cloud — Anthropic expanding its own model distribution channels;
- 06-30 · LongCat Owl Alpha #1 on OpenRouter — real penetration of a Chinese model into a neutral Western channel.
- Hermes Agent #1 — an open-source coding agent. LongCat is strongest in the open-source agent ecosystem;
- Claude Code #2 — Anthropic's own coding agent. This is an interesting signal: Anthropic users are proactively switching the backend model to LongCat Owl Alpha inside Claude Code. That implies Owl Alpha is at least comparable to Claude Opus 4.6 on coding tasks — otherwise users wouldn't switch;
- OpenClaw #3 — a multimodal agent, showing it is also usable in multimodal scenarios.
- The exact model of the 50,000 domestic ASICs is undisclosed — Huawei Ascend, Cambricon, or a newer domestic flagship? This determines how real the 'compute autonomy' claim is;
- The composition of the 35T-token training data is undisclosed — English data? Synthetic data? Data quality directly affects real-world English performance;
- Activated parameters of the 1.6T MoE are undisclosed — if activation is very small (e.g., under 20B), '1.6T' is a marketing number and inference speed is no different from a 27B dense model;
- Whether OpenRouter's 10T tokens of usage is 'cumulative' or 'recent' is unclear — cumulative would be a long-term reputation metric; monthly/weekly would be a short-term hype metric;
- Emad Mostaque's 'Gemini / Opus 4.6 level' claim deserves a discount — his style is consistently hyperbolic; the real level may be between Sonnet 4.5 and Opus 4.5;
- Whether Anthropic will officially respond to users switching Claude Code's backend to LongCat — a signal worth watching for the ecosystem.
This is reportedly the first trillion-scale model trained entirely on domestic ASICs to reach 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. Despite matching GPT-4-class models on international benchmarks, their real usage on mainstream Western distribution channels remained a "minority option."
LongCat Owl Alpha is different. Its usage scenario is coding agents + coding CLIs — toolchains Western developers use daily (Hermes, Claude Code, OpenClaw). Holding Top 3 positions inside a Western developer workflow is a genuine validation of the domestic model ecosystem's capability.
The "Gemini / Opus 4.6 level" claim deserves a discount — Emad's takes tend to be aggressive. But 10 trillion tokens of real usage on OpenRouter needs no discount: thousands of Western developers are using this model to write code every day.
What does "trillion-scale model trained on domestic ASICs" mean?
Technical details: 50,000 ASICs training a 1.6T MoE on 35T tokens. No GPUs involved.
Either way, this validates that in LLM training, the effectiveness of 'compute blockades' is decaying fast.
OpenRouter's role as a neutral platform.
OpenRouter is an LLM API aggregation platform that started in 2024 — it plugs in OpenAI, Anthropic, Google, Mistral, Meta, and various Chinese models behind a unified OpenAI-compatible API. Developers switch models with one key.
LongCat Owl Alpha reaching #1 on OpenRouter means:
Relationship to earlier related signals:
Together, these four point in one direction: the 'LLM gateway' is becoming the new 'app store' — whoever controls this distribution layer controls the channels of the AI era. LongCat Owl Alpha is the first Chinese model to reach the top of a neutral Western distribution channel.
Specifics of the use cases.
Why It Matters
1. "Trillion-scale model trained on domestic ASICs" is a key 2026 milestone for AI infrastructure self-sufficiency. Far more significant than "yet another Chinese LLM release" — it validates that compute blockades are losing effectiveness. 2. "A Chinese model at the top of a neutral Western distribution channel" is a new paradigm for going global. Previously, Chinese LLMs went overseas via "overseas subsidiaries + overseas partners"; now neutral platforms like OpenRouter offer a more direct route. 3. Model replaceability in the coding-agent era is real. Developers willingly switching Claude Code's backend from Opus to LongCat Owl Alpha shows model choice in coding agents is driven by actual results, not brand loyalty — a stress test for every model company. 4. Meituan's AI strategy has been severely underrated. Beyond its core consumer (delivery, flash sales) and merchant SaaS businesses, Meituan's AI investment has been underestimated for two years. A 1.6T MoE model trained on domestic ASICs means its AI capability is no longer bound to any external vendor. 5. OpenRouter's channel value may matter more than the model itself. Whoever becomes the 'app store' of the AI era controls distribution for the next decade.