---
title: "The 194 Chips in Meituan's Launch Post Never Trained Anything"
date: 2026-11-17
category: Foundation Models
site: NeuroAI
canonical: https://neuroai.site/a/na-model-longcat-domestic-1117
language: en
---

# The 194 Chips in Meituan's Launch Post Never Trained Anything

> The open-source announcement from Meituan (3690.HK) gives hardware one sentence: a domestic AI accelerator with under 80GB of memory per card. No model.

The open-source announcement from Meituan (3690.HK) gives hardware one sentence: a domestic AI accelerator with under 80GB of memory per card.

No model. No vendor. No card count.

Five days later The Decoder filled it in differently: training used domestic chips throughout, and the demo deployment ran on 194 Ascend cards. Chinese outlets passed it on once more, and it became "trained end to end on domestic chips, 194 cards."

In Meituan's materials, two separate claims.

## Where 194 comes from

It comes from the deployment note shipped with the weights, not a training log.

The Decoder's November 15 report: the self-hosting example is a 194-card Ascend Atlas A2 cluster, with 2 cards for image processing, 64 for prompt ingestion and 128 for response generation. The three add to exactly 194.

That is how inference splits, not training.

## Memory first, training second

1.6 trillion parameters at 4bit is half a byte each, about 800GB. Over the 128 generation cards that is 6.25GB each. With 64GB on an Atlas A2, more than nine-tenths is left for KV cache and activations.

That is why 194 cards suffice: not a small model, but weights cut to one-eighth. At 16bit the same model is 3.2TB, 25GB per card, and the room for KV cache falls from 58GB to 39GB. Concurrency drops by a third.

KV cache is the real cost of long context: a 1-million-token window adds a cache per concurrent session, held through generation, which is why 128 cards are set aside.

LongCat-2.0 shipped on June 30, 2026, when Meituan disclosed a training cluster of more than 50,000 domestic AI chips.

50,000 divided by 194 is 258.

Training and deployment differ by two orders of magnitude; putting 194 on the training side means one card did the work of 258. The accurate sentence is narrow: Meituan says training used domestic accelerators without naming them; the deployment example uses 194 Ascend Atlas A2 cards and is not called training. "No Nvidia anywhere" is The Decoder's phrasing, not a confirmation from Meituan.

## The 80GB wall

Section 5 of the blog, "domestic compute," names the constraint: under 80GB per card, long-sequence training and fine-tuning both stall.

The fix is sharding. FSDP double sharding splits parameters and gradients, flattened buffers lower peak memory, hybrid vocabulary parallelism takes another bite. Topology-aware communication keeps parameter AllGather, expert-token AllToAll and KV AllGather inside the supernode where possible, and MoonEP replicates experts only when imbalance crosses a threshold. Per-card throughput on long-sequence SFT ends up 86% above a DDP baseline.

## 4bit is not a zip file

W4A4 means weights and activations both at 4bit, computed on native INT4 units. A zip file has to be unpacked; W4A4 computes in 4bit directly, saving bandwidth and compute rather than disk.

Meituan's numbers: prefill throughput up 43.3%, decode up 42.77%, first-token latency down 15.91%, time per output token down 19.72%. Accuracy is "comparable to pre-quantization."

Self-reported, not yet reproduced.

## Vision, and a disconnected network

Native multimodality here is not a bolted-on encoder. Visual features go through residual vector quantization into multi-level discrete tokens, each level adding detail, and enter the same backbone as text. Codebook and backbone train jointly. Two benefits: visual tokens can be precomputed offline and cached, and distributed training balances better.

Images then consume context too.

Long-horizon agents learn shortcuts faster than solutions. Meituan lists 3 constraints: the environment is offline; git history and file caches are cleared; another model judges every trajectory. Repeated reasoning, repeated tool calls and overthinking carry penalties. Training data includes real failure trajectories, rewritten as reproducible tasks.

The constraints answer a specific failure: the model searching git log instead of fixing the bug.

## Read the scores sideways

On Terminal Bench v2.1, LongCat-2.5 scores 88.4, against 90.6 for DeepSeek-V4.1-Flash and 89.1 for Claude Opus 5: 2.2 points back, third.

On the harder v4.0 it scores 38.5, against 31.2 for DeepSeek, 41.8 for GLM-5.3 and 53.9 for Opus 5 — 23.4% above DeepSeek, 28.6% below Opus.

On DeepSWE, which fixes bugs in real repositories, it rose from 9.7 to 64.6, against 74.2 for DeepSeek and 72.7 for GPT-5.6 Sol. A 5.7-fold gain, still 13% behind.

A footnote says starred scores come from other vendors' reports, unstarred ones Meituan ran; few rows compare directly.

One Chinese write-up called it "second only to Claude Opus 5" on v4.0. GLM-5.3 sits at 41.8, above its 38.5; someone read two rows.

## What MIT actually grants

MIT grants commercial use, further training and redistribution.

It withholds training data, chip-hours, the evaluation method and an independent technical report. A blog post and weights on Hugging Face are what shipped. The weights and a deployment path are usable; the reproduction path is incomplete.

Two repositories: the main one and an expert-layer 4bit build. The second is the runnable one; at 16bit, 1.6 trillion is not a casual download. The licence is MIT.

## Why a delivery company does this

Meituan runs one of the world's largest dispatch systems daily: riders, merchants, routes, pricing, all long-horizon decisions. That load has the shape of a coding agent — many steps, tools, a real world to land in.

The model is a cost line, not a revenue line. Train it yourself if you can, open it if you can: open source buys community help on kernels, chip adaptation and bug hunting, the 3 things domestic hardware lacks most.

The same week, Alibaba locked 2.4 trillion behind an API and Meituan released 1.6 trillion. Neither is wrong. The books differ.

"Domestic AI accelerator" is honest and leaves room: it admits the constraint, and does not say whose chips or how many trained the model. The one blank that can be filled is 194, and it describes inference. Next time you see "trained on domestic chips," ask: training, or inference?

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

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-longcat-domestic-1117
Free to quote with attribution and a link to the original.
