NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

1.6 Trillion Parameters, Not One Nvidia Card

LongCat-2.5 holds 1.6 trillion parameters. Each token activates about 48 billion of them.

2026-11-22 · 4810 words · NeuroAI
1.6 Trillion Parameters, Not One Nvidia Card

LongCat-2.5 holds 1.6 trillion parameters. Each token activates about 48 billion of them. Divide one number by the other and you get 33.3, which means a token walks through roughly 3% of the network and leaves the rest untouched.

The idle majority is the design, not a side effect.

Meituan released the model in November under an MIT license. It is natively multimodal, taking text, image and video as input, with a 1M token context window. The part that matters for the deployment story is where the weights were made: the whole pipeline ran on domestic accelerator cards, according to the company's engineering documentation, and a single card carries under 80GB of memory.

Sparsity saves arithmetic, not memory. All 1.6 trillion parameters still have to sit somewhere, and if one card cannot hold them, they get split. Split weights must be synchronized across cards, and synchronization costs latency. That constraint produces the most concrete artifact in the release.

194 cards, itemized

The self-hosting guide specifies 194 Huawei Ascend Atlas A2 cards: 2 for image processing, 64 for prompt ingestion, 128 for response generation. The 3 numbers add to 194 exactly, with nothing held back, and the guide contains no Nvidia option anywhere in it.

A bill of materials, not a slogan.

The 1:2 split between reading and writing follows the shape of inference. A question is read once. An answer is emitted token by token, and the emitting stage is the expensive one. With a 1M token context the reading stage grows too: 64 cards must ingest the whole prefix, and the intermediate state has to stay resident afterward. Longer reads make the writing side more expensive.

1.6 trillion divided by 194 is 8.25 billion parameters per card on average. That division ignores activation and parallelism, so it is not a memory plan. It only pins down what the headline means. No single card does the work. 194 cards each do a slice, and none of them is Nvidia.

194 cards is also the threshold for self-hosting. This is not a workstation. Cards must be interconnected, memory pooled, requests queued separately by direction. An open license distributes the model. It does not distribute the machine room.

What the benchmark says, and who said it

The scores are vendor-reported. Terminal Bench v2.1: 88.4. Claude Opus 5 scores 89.1 and GPT-5.6 Sol scores 88.8 on the same benchmark. BrowseComp: 91.9. Put the three Terminal Bench figures side by side and one claim survives. LongCat-2.5 is in the leading group, 0.7 points off the top, not 7.

A vendor-reported score is a number waiting for a rerun. That is not the same as a doubtful number, but no third party has reproduced it, and both the benchmark selection and the reporting were done by the publisher.

4 bits instead of 8

The more informative figure is quantization. Against an 8-bit baseline, W4A4 raises prefill throughput by 43.3% and decode throughput by 42.77%. The two gains differ by 0.53 percentage points, which is close to symmetric: reading and writing improved together, so the compression did not favor one stage over the other.

4 bits is half of 8. Weight volume halves, the parameters one card can hold double, and the bytes that must move between cards halve as well. Bandwidth is a cost line too.

Training and serving ledgers

LongCat-2.0, the predecessor, was trained in late June on a cluster of more than 50,000 domestic chips. Late June to November is under 5 months. The two numbers cannot be subtracted from each other. 50,000 is a training cluster; 194 is an inference reference configuration. One builds the model, the other serves it. Training optimizes for parallelism and fault tolerance, inference for latency and throughput.

2 ledgers, 2 rulers.

A training run has tens of thousands of cards stepping in lockstep, and an idle minute is money. A serving deployment runs continuously, loaded at peak and empty off peak. Comparing the two configurations in one table measures the difference between two kinds of work, not progress. What can be compared is time: under 5 months from a 50,000-chip cluster to a 194-card deployment sheet.

A sheet is easier to copy than a cluster.

One salon, 7%

The application evidence is small. Meituan's Miaoshou merchant AI workspace runs on the LongCat family. One Dushang Hair Salon store tried it in August and recorded 7% month-over-month GTV growth in September. One store, one month, one case. It does not measure the model. It shows the model is wired into a real store's operations and left behind a figure that can be recomputed.

The 7% is month over month, not year over year, for one month, at one store. Drop any one of those qualifiers and the number inflates.

Proven feasible, not yet repeatable

Cai Xunliang, who leads Meituan's LongCat large model effort, said the team progressively solved deterministic computation on neural processing units and improved compute utilization, which he described as verifying the feasibility of large-scale domestic training. The operative word is feasibility. Feasibility is a proof: it holds once it has been done. Repeatability is an engineering claim: someone else has to do it too.

Deterministic computation is the concrete part. The same data and the same parallel split must produce the same result on a rerun, or a problem on a large cluster cannot be reproduced and tuning has no baseline. That is what makes 194 cards with fixed roles worth reading. 2 for images, 64 for input, 128 for output. Anyone who wants to reproduce the deployment can buy exactly that.

Training still happens on clusters of 50,000 chips. That part did not become a sheet.

The sheet is page one.

*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*

Related coverage

Cite this article

NeuroAI (脑机网), "1.6 Trillion Parameters, Not One Nvidia Card," 22 November 2026, https://neuroai.site/a/na-model-meituan-longcat-domestic-1122

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Foundation Models” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories