Meituan releasing a trillion-parameter model is not surprising. What surprised everyone was how it released one: the model had already been running anonymously on OpenRouter for two months, climbing to top-three global traffic and second place in the Claude Code ecosystem, before the company stepped forward to say — that owl was us.
The second surprise came in a single sentence: from training through inference, the model never used one NVIDIA card.
Chinese media almost universally framed this as “domestic chips overtaking on the curve,” a satisfying narrative and a dangerous one, because it compresses a structural story that deserves cold analysis into an emotional label. LongCat-2.0’s real significance is not the binary question of whether domestic silicon works. It is three facts the emotion drowned out: domestic compute has graduated from “can it” to “is it stable”; the scaling dividend of MoE architecture is hitting a measurable ceiling; and — most underpriced of all — real-world scenario data is becoming a scarcer competitive moat than model architecture itself.
Five Million Cards Worth of Context: Why “From Scratch” Matters
To weigh the compute achievement, separate two milestones the news cycle blurred together. In early June, a Shenzhen institute working with Harbin Institute of Technology and Huawei completed full-parameter post-training of DeepSeek-V4-Pro on an Ascend 910C cluster — fifteen hundred iterations, no interruptions, MFU above 30 percent. A real accomplishment: it proved domestic cards can fine-tune a trillion-parameter model.
Meituan did the harder thing. On roughly 50,000 Ascend 910C cards, it pretrained a 1.6-trillion-parameter MoE model from a random initialization, across more than 30 trillion tokens, driving the monthly average daily failure rate down over 70 percent, lifting training MFU 1.5-fold, and finishing with no rollbacks and no unrecoverable loss spikes. The analogy one developer offered: post-training is renovating a finished house with domestic tools; LongCat-2.0 is pouring the foundation through move-in — and the residents report it lives well.
The single-card compute of the 910C reportedly approaches an H100. The real choke point was never raw FLOPS; it is the software stack — the CANN and HCCL ecosystem’s maturity gap against CUDA. Meituan’s heavy lifting went into low-level operator adaptation and communication optimization: key operator efficiency up 14 percent, MFU past 30 percent, all resting on HCCL exception handling and automatic fault recovery. The significance is not “domestic chips beat NVIDIA.” It is industrial-grade proof that a domestic cluster can carry the complete loop — pretraining to inference deployment — of a trillion-parameter model without foreign hardware. For an industry living under export controls, that is a phase change from “can it” to “is it stable.”
97% Sparsity: The Expert-Stacking Race Is Over
If compute is the external narrative, architecture is where Meituan’s technical judgment shows. One line from the official blog outweighs the 1.6-trillion headline: the model’s MoE sparsity, excluding N-gram embeddings, has reached about 97 percent, and adding another 135B expert parameters would yield negligible gains.
For calibration: DeepSeek-V3 is 671B total parameters activating 37B, about 94 percent sparse. V4-Pro is 1.6T activating 49B, about 97 percent. LongCat-2.0: also 97 percent. Three independent teams, three architecture processes, one asymptote. The implication is stark. For two years everyone raced to stack experts — 560B to 671B to 1T to 1.6T — each generation convinced that one more zero would unlock proportionally more capability. Meituan’s data says: at 97 percent, additional experts are statistically invisible. Double the expert pool to 3.2T and each token still activates the same ~48B; you have doubled hardware and memory costs to compute exactly what you computed before.
Which is precisely why LongCat-2.0 is smarter than its own spec sheet: it stopped inflating the expert pool and redirected the effort to attention mechanisms, context efficiency, and post-training data — LongCat Sparse Attention, N-gram embeddings, the Muon optimizer, 6D parallel training, multi-objective post-training. The whole combo says one thing: extract more from a fixed expert pool rather than enlarging it. If this plateau holds, the next phase of competition shifts from “whose model is bigger” to “whose architecture is more efficient, whose data is better, whose post-training is sharper.”
Two Months Anonymous: Developers Voted With Tokens
The most underrated aspect of this release is that the launch strategy itself is a signal. Meituan skipped the launch event, the press tour, the curated benchmarks. It listed the model on OpenRouter as “Owl Alpha” — free, fast, eerily good at code completion — and let global developers use it for weeks with no idea who built it, until on-chain messages unmasked the owl.
The tally at reveal: total traffic in the global top three; number one worldwide in monthly calls on the Hermes harness; number two in Claude Code, behind only Claude Opus itself. That last position deserves a long look. Claude Opus is Anthropic’s flagship, the de facto standard of agentic coding, and Claude Code is Anthropic’s own garden. A Chinese model listed anonymously finishing second in that garden means that on the things agentic coding actually demands — comprehending repository structure, handling multi-file dependencies, producing engineering plans you can land — it hit real developers where they live. This was not leaderboard traffic; it was billable token consumption from users who did not know or care where the model came from.
Overseas commenters keep asking how a food-delivery company built a top-tier model. The answer hides inside the word “delivery.” Meituan processes tens of millions of daily orders and orchestrates millions of riders in real time; that dispatch complexity is itself one of the world’s best training grounds for agentic behavior — goal decomposition, dynamic replanning, tool orchestration under uncertainty. The exact split between base-model capability and post-training data has not been published, but the direction is hard to dispute: real-scenario data is becoming a scarcer moat than architecture. This is also why Anthropic’s Fable 5, for all its raw power, reportedly degrades quickly on real coding tasks — capability without scenario depth is a sprinter without a race plan.
A Framework: Three Validations, One Industry Signal
Layer the findings and you get a compact test any model release can be put to:
Supply-chain validation. Can domestic compute carry a trillion-parameter model’s full lifecycle? Yes — and stably. That converts “can it” into “is it stable” for every Chinese AI firm under export pressure.
Architecture validation. How much headroom is left in MoE scaling? Roughly none past 97 percent sparsity; the contest moves to efficiency and data quality.
Scenario validation. Does real business data convert into model competitiveness? Demonstrably — an agent trained in delivery dispatch out-drew most generalists in agentic coding, on their own terms.
The same tripod explains the neighbors. DeepSeek stays dangerous through extreme pricing — V4-Flash at $0.42 per million tokens builds its own developer scenario. Zhipu’s GLM-5.2, half LongCat’s size at 744B/40B, holds a top domestic coding tier purely on post-training quality — architecture validation in action. And the framework’s edge cases are instructive: raw capability without scenario grounding is exactly where Fable 5 stumbles.
What You Should Do With This
If you are an AI founder: stop chasing parameter narratives; 97 percent sparsity is a published ceiling. The next openings are in data quality and scenario depth. If you own proprietary workflow data, your model can outcompete generalists that benchmark higher.
If you are a technical decision-maker: the “can it” question for domestic compute is settled; price in “is it stable” and “is it cheap.” Note the caveat — Meituan spent three years on operator and communication optimization that not everyone can replicate, so lean on mature domestic platforms unless you have that bench strength.
If you are a developer: actually try the model. Pricing is aggressive — cache hits are free, and promotional rates run $0.30 per million input tokens and $1.20 per million output tokens, undercutting DeepSeek-V4-Flash — and its agentic coding has already been validated by the world’s most honest jury, usage.
If you allocate capital: weight companies with proprietary scenario data over companies stacking parameters. When architecture scaling stalls, data quality is the new scarcity — and a delivery company just proved the moat is real.
Sources: Meituan LongCat official blog; 36Kr; QbitAI; VentureBeat; Geopolitechs; Zhihu technical teardowns.
