---
title: "The company that delivers your dinner just open-sourced a 1.6-trillion-parameter model"
date: 2026-11-16
category: Foundation Models
site: NeuroAI
canonical: https://neuroai.site/a/na-model-longcat-25-mit-open-weights
language: en
---

# The company that delivers your dinner just open-sourced a 1.6-trillion-parameter model

> Meituan released LongCat-2.5 under the MIT licence on November 9 — 1.6 trillion parameters, a million-token context, and a 3.56TB download most teams cannot serve.

The download is 3.56 terabytes — 208 files, per the model card — and when the last one lands you own — legally, permanently, with no strings attached — a trillion-parameter model you can run without asking anyone's permission.

Meituan, the company most people outside China know as a food-delivery app (外卖), pushed the full weights of LongCat-2.5 (龙猫 2.5) to Hugging Face on November 9, 2026. The licence is plain MIT. No sign-up form, no revenue threshold, no list of countries where you are not allowed to use it.

The parameter count is the headline. The licence is the story.

## What was actually published

LongCat-2.5 is a native multimodal mixture-of-experts (混合专家) model: 1.6 trillion total parameters, roughly 48 billion of them active on any single token, per the model card. Because only a sliver of the network fires per token, per-token compute resembles a mid-sized dense model rather than a trillion-parameter one — which is the only reason a 1.6T system can be served at all.

The rest of the spec sheet:

- A **one-million-token** context window, with long-sequence training during the mid-training stage.

- Vision handled by converting image features into discrete tokens in the same space as text, via a component the model card calls LongCat-Next. No bolted-on vision encoder.

- A **hybrid thinking switch**, toggled per request through chat_template_kwargs, so a caller can buy deliberation or skip it.

- An OpenAI-compatible API, tool calling included.

- Two weight formats: BF16 and a **W4A4** build that quantises expert weights and activations to four bits for native INT4 compute.

One wrinkle before deployment: Hugging Face's metadata has displayed a 1.8-trillion figure against Meituan's stated 1.6 trillion. VentureBeat flagged the mismatch and advised sizing from the files rather than the label.

## Why a delivery company ships trillion-parameter models

Meituan's LongCat team has been building models for a while; LongCat-2.5 is the successor to LongCat-2.0 and the first in the line to take images natively.

The stated target is long-horizon agent (智能体) work: large codebase development, whole-project builds, visual coding, 3D tasks, professional office documents. Read that list against Meituan's business — dispatch, merchant tooling, instant retail — and the logic tightens. These are workflows made of hundreds of small, boring, verifiable steps, executed by software that has to keep running when nobody is watching. That is the shape of task where an agent either compounds value or compounds damage.

Meituan's technical write-up is unusually candid about how the gains were produced. Real agent failure trajectories were converted into reproducible training tasks, so the model learns from mistakes agents actually make. Reinforcement learning was made asynchronous, with partial rollouts saved and resumed so a half-finished task need not restart. Training environments were cut off from the network, git history and file caches cleared, and a judge model supervised trajectories — anti-shortcut plumbing, because an agent that can read the answer key will read the answer key.

Meituan also says the whole training run happened on domestic AI accelerators (国产算力). That claim has not been independently audited, but it is the part Chinese labs care most about right now.

## The benchmark table, and its asterisks

The model card is generous with numbers, and Meituan marks which ones are whose.

On **Terminal Bench v2.1**, LongCat-2.5 reports 88.4, up from 70.8 for LongCat-2.0. The same table lists Claude Opus 5 at 89.1 and DeepSeek-V4.1-Flash at 90.6, both cited from public reports. Read carefully: LongCat is *close to* the leaders on that suite, not ahead of them.

Other reported results: DeepSWE v1.1 at 64.6 (from 9.7), SWE Atlas 58.3 (from 33.1), BrowseComp 91.9, AutomationBench 51.8.

Then the number that matters most. On **Terminal Bench v4.0** — the harder, newer task set — LongCat-2.5 scores 38.5. Claude Opus 5 is listed at 53.9 on the same page. That is a 15-point gap on the suite that most resembles open-ended real work.

Meituan states plainly that unmarked scores are measured in-house and are indicative only, and that the model has not been comprehensively evaluated for every downstream use. As of this writing there is no independent evaluation to check any of it against.

## MIT is doing more work than the parameters

November 2026 turned into an accidental natural experiment in open-weight (开源权重) licensing, and the three big releases differ less in capability than in legal texture.

Kimi K3.1 Pro shipped under a modified MIT that, per third-party readings of the licence file, requires a separate agreement above a revenue threshold for model-as-a-service use. Baidu's ERNIE 5.5 Flash Base uses a custom licence that, by the same reckoning, carves out the EU, UK and South Korea and ties commercial use to revenue bands. LongCat-2.5's LICENSE file is about a kilobyte of plain MIT with a 2026 Meituan copyright line.

For a startup, that difference is not academic: MIT removes a whole category of questions — no revenue measurement, no jurisdiction map, no negotiation.

## Who can actually run it

Almost nobody can run it at full precision on hardware they already own — and that is fine.

Meituan's documented serving example uses **194 Atlas A2 accelerators**: two for encoding, 64 for prefill, 128 for decode, driven through SGLang-FluentLLM. That is a cluster, not a workstation. The company reports the W4A4 build delivering 43.3% higher prefill and 42.77% higher decode throughput than a W8A8 baseline, and 92% higher decode with KV-cache quantisation, per the model card. The tested path targets neural processing units; a GPU route through vLLM was still community work at the time of release.

Which is why the hosted API is the realistic entry point: **US$0.30 (≈HK$2.3) per million uncached input tokens, US$1.20 (≈HK$9.3) per million output tokens, and US$0.006 (≈HK$0.05) per million cached input tokens**, on a limited-time rate card.

## Honest limitations

- Every performance number above is Meituan's own or a citation of someone else's public report. No independent replication exists yet, and the card says so.

- The 1.6T-versus-1.8T metadata discrepancy is unresolved in public; size your deployment from files, not from a blog post.

- The published serving recipe is NPU-specific. If your infrastructure is NVIDIA, you are waiting on community work, not following a guide.

- Terminal Bench v4.0 is where the gap actually lives. A model can look frontier on v2.1 and mid-table on v4.0 without anyone lying.

- This article did not test the model or inspect its training data.

## What readers can do now

- **Rent before you download.** At US$0.30 per million input tokens, a week of real evaluation costs less than the electricity for the download. Run your own tasks through the hosted API, then decide whether self-hosting is worth a cluster.

- **Read the licence file, not the launch post.** The three biggest releases of the month all ship weights; only one is plain MIT. If you are building a product on top, that sentence is worth ten minutes.

- **Score it on the suite that hurts.** If a vendor quotes a Terminal Bench number, ask which version. LongCat-2.5 itself scores 88.4 on v2.1 and 38.5 on v4.0 — the two suites are not comparable, and the drop says more about the benchmark than about the model.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-longcat-25-mit-open-weights
Free to quote with attribution and a link to the original.
