Two weeks before Meituan shipped LongCat-2.0, a widely shared essay asked the anxious question of Chinese AI in one sentence: without access to NVIDIA’s best chips, can anyone still build a world-class model? For a year and a half that anxiety had calcified into consensus — from the A100 sales ban through H20 restrictions, from CUDA’s apparent irreplaceability to the engineering moats around giant training clusters, everyone kept asking the same thing: does domestic compute actually work?
On June 30, Meituan answered with a model that renders the question obsolete. LongCat-2.0: 1.6 trillion parameters (about 48 billion average activation), mixture-of-experts architecture, one-million-token context. Those numbers alone would not turn heads on today’s leaderboard — DeepSeek-V4-Pro posts 1.6T/49B/1M and even edges ahead on some benchmarks. But LongCat-2.0 carries a label no competitor can claim: from pretraining through inference, it never touched a single NVIDIA or AMD GPU. Roughly 50,000 domestic accelerators (community consensus points to Huawei’s Ascend 910C), pretrained from zero on 30 trillion tokens, stable to the end — no rollbacks, no unrecoverable loss spikes. Not a lab demo. A production run on the real clusters of China’s largest food-delivery company.
Here is the honest part, though: if LongCat-2.0 only proved “domestic chips can train trillion-parameter models,” it would be a political narrative rather than a technical story. What makes it genuinely important is the second ceiling it exposed — one that applies to the entire industry, Silicon Valley included.
The First Wall: An Engineering Miracle of From-Scratch Pretraining
Word choice matters here: “pretraining from scratch” and “post-training on domestic chips” are different sports. In early June, a team in Shenzhen partnered with Harbin Institute of Technology and Huawei to run full-parameter post-training of the existing DeepSeek-V4-Pro weights on an Ascend 910C cluster — 1,500-plus stable steps, a genuine engineering feat that proved domestic cards can fine-tune a trillion-parameter model. But fine-tuning is renovating a house whose foundation someone else poured.
From-scratch pretraining is pouring your own foundation, then building the tower. A 1.6-trillion-parameter MoE model on 50,000 domestic cards confronts you with the whole inventory: single-card memory far below the H-series (the 910C carries roughly 64GB of HBM against H100’s 80GB), interconnect bandwidth behind NVIDIA’s NVLink/NCCL pairing, a CANN/HCCL software stack with large maturity gaps against CUDA, and failure rates that climb exponentially as the cluster scales.
Meituan’s engineering answers, from the public disclosures, read like a siege campaign. HCCL exception handling plus automatic fault recovery cut the monthly average daily failure rate by more than 70 percent. Fully deterministic self-developed operators guarantee bitwise-consistent results, so a rerun never diverges. The teams squeezed pipeline scheduling, memory management, and per-core operator tuning until training MFU rose 1.5x past 30 percent, with steady-state throughput exceeding one trillion tokens per day. One engineer who has lived through similar work described domestic pretraining as building the same tower as an NVIDIA cluster with smaller shovels and slower conveyors — and no inherited blueprints, because the CUDA playbook simply does not transfer. LongCat’s team finished the tower, and the occupancy reviews — global top-three total traffic on OpenRouter under the anonymous alias Owl Alpha, second only to Claude Opus inside Claude Code itself — suggest the tower is comfortable to live in.
The wall is down. But its real meaning is not “domestic substitution succeeded.” It is that Chinese AI’s core question has shifted from “can we?” to “should we?” — no longer a question of technical possibility but of commercial allocation: is it worth investing the engineering to adapt and optimize? The answer is now self-evident, and one corollary deserves more attention than it got: non-cloud companies running self-built compute are the best soil for domestic-chip adaptation — unburdened by public-cloud CUDA stack lock-in and millions of person-months of migration cost, they can choose hardware from zero like a startup does.
The Second Wall: MoE Sparsity at 97% Ends the Stacking Race
If the compute story were all LongCat-2.0 had, it would be filed under “milestone in domestic substitution” and scrolled past. But Meituan’s technical blog buried one sentence that matters more than the chip story, and it deserves to be quoted verbatim:
The model’s MoE sparsity, excluding N-gram Embedding, has reached about 97%; further expanding expert parameters by 135B yields performance gains that are negligible.
Unpack it. The core logic of MoE is to split a model into many “experts” and activate only a small slice per token, so total parameters can grow while per-token compute stays capped. DeepSeek-V3: 671B total, 37B active, roughly 94% sparsity. V4-Pro: 1.6T over 49B, about 97%. LongCat-2.0: also 97%. And Meituan’s data says that at 97%, adding experts produces gains too small to measure.
A brutal counterfactual makes the point. Suppose LongCat doubled its expert pool to 3.2 trillion total parameters. At 97% sparsity, every token still activates the same ~48B. You pay double for hardware and memory to run a forward pass indistinguishable from the current one. That is not scaling; that is purchasing the privilege of a bigger number on a model card.
This is an asymptote problem, not a tuning problem. For two years the industry’s main melody was stack the experts: 560B, then 671B, then 1T, then 1.6T, with total parameter count as the marketing story on every funding deck. Meituan’s number declares that road’s end — and DeepSeek-V4-Pro landing on the same 97% independently means this is not one team’s coincidence; it is a structural signal about the entire technical route. When “bigger” stops buying “better,” competition pivots to the dimensions Meituan itself pivoted to: attention efficiency through LongCat Sparse Attention, commercially viable million-token context, post-training data quality via multi-objective MOPD, and inference cost control. That is the post-MoE battlefield, and it rewards engineering depth rather than parameter volume.
Two Walls, One Rewritten Narrative
Put the results side by side and LongCat-2.0 has broken through two ceilings at once: the compute-autonomy wall — from “can domestic chips run inference” through “can they train trillion-parameter models” to “can they do it profitably at scale” — and the architecture-efficiency wall, where MoE competition moves from “whose model is bigger” to “whose model is smarter.”
Before June 30, you could hear it at any conference: domestic chips handle inference but cannot train. That consensus is now shattered — not by a chip vendor’s slide deck, but by a company that serves hundreds of millions of daily users spending real engineering years to prove otherwise. The cost matters as context: Meituan’s AI investment is plausibly the largest of any Chinese enterprise outside the cloud giants, spanning chip design (Moore Threads, MetaX, UNISOC) through embodied AI (Unitree, Galaxy General). This was a strategic bet on a three-year clock, and the payoff arrived on schedule. Meanwhile the second wall’s fall ends the parameter-count era for everyone: if 97% sparsity is the asymptote, additional total parameters buy marketing budget, not capability. For DeepSeek, Zhipu, Alibaba, ByteDance, and the rest, the past two years’ arms race is over — not because larger models became impossible, but because larger models stopped mattering. The next round is efficiency optimization, scenario fit, and cost control, decided by engineering depth.
Three Groups, Three Signals
If you build foundation models — stop competing on total parameter count. Within a fixed activation budget of 30-50B, the real battles are post-training quality, sparse-attention efficiency, inference cost, and real-world agent performance. Move the expert-expansion budget to those fronts now, because Meituan and DeepSeek have both published the ceiling you would be racing toward.
If you make domestic chips — LongCat’s engineering feedback is rare, genuine production data. Single-card memory, interconnect bandwidth, operator coverage, cluster stability: these are the places model teams actually bleed. Concentrate resources on those pain points instead of single-card benchmark records; the buyers read failure rates, not launch decks.
If you build AI applications — LongCat-2.0 is open-sourced under MIT, with second-place global traffic in agentic coding on OpenRouter. You can run a trillion-parameter model validated on real code tasks, at aggressive cost (Meituan prices it around 9.9 RMB per 50 million tokens). For teams building coding tools, the combination of an open-weights model plus low-cost domestic inference is not a national-pride headline; it is infrastructure arbitrage, available today.
Sources: Meituan LongCat official technical blog; IT Home; 36Kr; QbitAI; community teardowns and Geopolitechs coverage. Weights: Hugging Face, meituan-longcat/LongCat-2.0.
