Meituan's LongCat-2.5: 1.6T parameters on 194 Huawei chips

Meituan put the weights for LongCat-2.5 on Hugging Face this week under the MIT licence: 1.6 trillion parameters, about 48 billion of them active per token, one-million-token context. The interesting part is not the benchmark table. It is the deployment recipe attached to it — a published way to serve a model that size on accelerators that are not Nvidia's.

What actually shipped

The model was announced on November 2 and the weights followed on November 9, in two repositories: a full BF16 release around 3.56 TB and a 4-bit version around 1.34 TB, per Meituan's LongCat team. MIT means commercial use, modification and redistribution without asking, which is a looser licence than most Western "open" releases.

Architecturally it is a mixture-of-experts model built for long, multi-step agent work rather than chat. Its visual handling is the unusual bit: images are converted into discrete tokens at native resolution using residual vector quantisation, the same trick Meituan described in LongCat-Next, so the main model processes pictures and text in one vocabulary instead of bolting on a separate image encoder. Because those tokens can be precomputed and cached offline, the vision encoder drops out of the training graph entirely — a real systems saving when you are packing million-token sequences.

It accepts text, image and video, answers in text, and has a thinking toggle. The API speaks both OpenAI and Anthropic formats, so it drops into existing agent harnesses without a rewrite. Promotional pricing runs at roughly €0.27 per million input tokens and €1.07 per million output tokens, according to The Decoder, which is cheap enough that price is not the constraint on trying it.

The 194 chips are a serving recipe, not a training claim

The framing that travelled fastest is "never touched an Nvidia GPU". Meituan does say training ran on domestic accelerators, and its engineering write-up gives one hardware detail: less than 80 GB of memory per card. It does not name the vendor or the model, and it publishes no chip-hours or cost.

The 194 figure that headlines have attached to that claim is inference. Meituan's self-hosting guide describes a demonstration cluster of 194 Huawei Ascend Atlas A2 accelerators split three ways — 2 for image processing, 64 for prompt ingestion, 128 for response generation — as the model card and deployment notes set out. The 50,000-chip cluster that gets quoted alongside it belonged to LongCat-2.0 back in June. Training a 1.6-trillion-parameter model on 194 cards is not a thing anyone is claiming, but the two numbers have been sitting in the same sentence all week.

The technical substance is elsewhere. Meituan quantised expert weights and activations to 4 bits (W4A4) and runs the matrix multiplies on native INT4 units, which is what lets a model this size fit on cards with under 80 GB and limited bandwidth. Against an 8-bit baseline it reports 43.3 percent higher prefill throughput, 42.77 percent higher decode throughput, 15.91 percent lower time to first token and 19.72 percent lower per-token latency, with grouped matrix kernels 1.73x faster. Add KV-cache quantisation and it claims 92 percent more decode throughput than the 8-bit equivalent.

That is the part worth taking seriously, and the part with an honest catch. The Chinese analysis site ic.work, reading the same primary material, points out that the training side never names a chip, that LongCat-2.5 was distilled from several unnamed teacher models, and that the inference framework's main branch still carries notes for testing on Nvidia H800 and H20 parts. It also walks the self-hosting path: CANN 9.0, a specific PyTorch and torch_npu pairing, and two custom operator packages you compile yourself, with no Nvidia W4A4 path provided at all.

So "no Nvidia" is accurate for the artefact you can download and the recipe you can follow, and unverified for the training run. Both of those are true at once, and only one of them is checkable by a reader.

What the benchmark table says, asterisks included

On Terminal Bench v2.1, the command-line agent benchmark, LongCat-2.5 scores 88.4. That is up from 70.8 for LongCat-2.0, and it lands within a point of Claude Opus 5 at 89.1, GPT-5.6 Sol at 88.8 and Kimi K3 at 88.3. DeepSeek-V4.1-Flash leads at 90.6.

The harder variant tells a different story. On Terminal Bench v4.0 the model scores 38.5 against Opus 5's 53.9, and on DeepSWE v1.1 — bug-fixing in real repositories — it manages 64.6 against DeepSeek's 74.2 and Opus 5's 68.8. BrowseComp, a web-research benchmark, is its best result at 91.9, the top of Meituan's table, run in Meituan's own harness.

Read the footnotes before quoting any of it. Meituan marks competitor scores with an asterisk where the number was lifted from that lab's own report or a public leaderboard rather than re-run. The three-run averages are its own; most of the comparisons are not. A model that improves 55 points on DeepSWE in one generation is doing something real, but the gap to the frontier widens exactly where the work gets least glamorous — long, hard, multi-hour terminal tasks.

Who wins and who loses

Huawei wins, indirectly and substantially. An MIT-licensed trillion-parameter model with a published 194-card serving layout and a 4-bit quantised twin is precisely what makes an accelerator ecosystem usable: without weights tuned for it, Ascend is a spec sheet. Meituan has now shipped two generations of that, and the W4A4 work is transferable engineering, not a one-off demo.

Nvidia loses a little of its lock-in argument, though not as much as the headlines suggest. China is still importing H200s for the labs that can get them — we covered Beijing letting ByteDance and Tencent import them — and DeepSeek's planned 160,000-chip Ascend cluster is where the real test of domestic silicon at scale sits. Meituan's contribution is the software layer that makes such clusters worth building.

The open-weights purist loses too, mildly. A 3.56 TB BF16 checkpoint served across 194 accelerators is not a model most teams will self-host; the realistic consumer is an API customer or a lab with a cluster. MIT is genuinely permissive, and the usable path is still someone else's data centre.

And Meituan gets what it actually wanted: adoption. At €0.27 per million input tokens it is not running a profitable token business; it is buying developers for the agent layer underneath its own products. For a food-delivery company competing on local services, owning a competent coding and computer-use model is cheaper than renting one from a rival.

What to watch

Independent re-runs, on the same harness, of Terminal Bench v4.0 and DeepSWE. Meituan's table is a manufacturer's specification; the open-weights community usually produces the second opinion within a month, and the 15-point gap to Opus 5 on the hard terminal benchmark is the number that will move first.

Second, whether anyone publishes an Nvidia W4A4 path for this model. It would prove the quantisation work is hardware-agnostic — and it would quietly retire the cleanest marketing line attached to it.

Would you rather run a model that fits your accelerators, or one that fits the leaderboard? Write to us via our contact page.