Skip to departures
Hpjadul
MAC END

Meituan's LongCat-2.5 Opens Weights for a 1.6T Agent Model Run on Huawei Chips

Departs
Crew
Jonathan Kemper
Passengers
625

Meituan's 1.6-trillion-parameter LongCat-2.5 ships open weights under MIT, with a self-hosting guide for a 194-chip Huawei Ascend cluster and no Nvidia path.

Meituan's LongCat-2.5 Opens Weights for a 1.6T Agent Model Run on Huawei Chips

Meituan has released the weights for LongCat-2.5, a 1.6-trillion-parameter model aimed at autonomous agents, and the most notable detail is the hardware. The model was trained entirely on Chinese chips, and its published deployment guide doesn't mention Nvidia GPUs at all.

Jonathan Kemper

The LongCat team at the food delivery company released the weights about a week after the model was unveiled. LongCat-2.5 targets agents that carry out long, multi-step tasks without supervision, with coding as the main use case. Meituan says it comes close to leading Western models on several agent benchmarks. Claude Opus 5 reached similar coding scores only a few months ago, so the gap between Chinese open-weight models and the US frontier is narrowing faster than many expected.

What the architecture actually does

LongCat-2.5 is a mixture-of-experts model. For each token, a router activates only a subset of specialized subnetworks, so the model carries 1.6 trillion parameters but uses roughly 48 billion of them per token. That split is what makes a model this size practical to serve. The API accepts up to one million tokens of input and generates responses up to 128,000 tokens long. It handles text, images and video, replies in text, and lets users switch reasoning on or off per request.

The new piece is vision. LongCat-2.5 builds on its predecessor, LongCat-2.0, and adds image understanding. Many multimodal systems bolt a separate image encoder onto a language model after training. LongCat-2.5 takes a different route: it converts images into discrete tokens at their native resolution and processes them in the same sequence as text, which is how most current LLMs handle multimodal input. The approach draws on the research model LongCat-Next and is meant to help with visual coding, design and 3D tasks. Native tokenization matters here because a model that sees layout and fine detail through the same token stream has a better chance of reasoning about a screenshot of a UI than one that receives a compressed summary of it.

LongCat-2.5 breaks images down at their native resolution into discrete tokens that the model processes alongside text. | Image: Meituan

How it was trained for agent work

The reinforcement learning setup is where the agent focus shows most clearly. Tasks were built from real software work: GitHub pull requests, plus recreated examples of agent failures. The harder problem is verification. An agent can look like it succeeded by finding the answer somewhere it shouldn't, such as a web search, the Git history or a cached build. To close those routes, the developers cut internet access in the training environments, removed Git history and cached files, and used a language model to check whether solutions relied on tricks rather than actual fixes. Anyone building agent training pipelines will recognize this as the hard part: reward signals are only as good as the sandbox that produces them.

The team also distilled capabilities from several specialized teacher models, covering agent behavior, reasoning and dialogue, into the final model. Distillation from narrow experts is a common way to pack several skills into one checkpoint, though Meituan's report doesn't break down how much each teacher contributed.

Image description
Overview of the LongCat-2.5 reinforcement learning pipeline with five sections, from task synthesis using GitHub repositories through a sandboxed environment with no web search, Git history, or build cache, to reward verification via tests and an LLM judge.
Meituan generates training tasks from real GitHub repositories and blocks shortcuts like web search and Git history in the sandbox. | Image: Meituan

Serving it on Huawei Ascend

Meituan quantizes the experts' weights and activations to 4-bit precision, a format its target chips can process natively. In its load tests, the 4-bit version delivers 43.3 percent higher throughput for prompt processing and 42.77 percent higher throughput for response generation than an 8-bit version. Meituan says accuracy stays on par with the uncompressed model. Those figures come from the company's own tests, so they're worth reproducing on your own workload before you plan capacity around them.

The 4-bit weights live in a separate Hugging Face repository. The self-hosting guide describes a sample cluster of 194 Huawei Ascend accelerators from the Atlas A2 series. Two handle image processing, 64 handle prompt ingestion and 128 handle response generation, which reflects the split between prefill and decode that serving stacks use to keep each stage busy. The guide contains no instructions for Nvidia GPUs. For a team that can't or won't buy Nvidia hardware, that's a concrete recipe. For everyone else, it's a sign of how far the Ascend software path has matured.

Where it trails the leaders

On Terminal Bench v2.1, which measures how well agents complete command-line tasks, LongCat-2.5 scores 88.4. Meituan's table puts Claude Opus 5 at 89.1, GPT-5.6 Sol at 88.8 and Deepseek-V4.1-Flash at the top with 90.6. The gaps are small, and this is the benchmark where LongCat looks most competitive.

The DeepSWE results show a larger jump. DeepSWE tests whether a model can fix bugs in real code repositories, and LongCat-2.5 rises from 9.7 points for its predecessor to 64.6. That puts it behind every model in the table except Claude Opus 4.8, with Deepseek-V4.1-Flash leading at 74.2 and GPT-5.6 Sol at 72.7. Claude Opus 5.5 and GPT-6.1 Sol are not in the comparison, so the picture is incomplete at the current frontier.

Bar charts comparing LongCat-2.5 and LongCat-2.0 with Claude Opus 5, Claude Opus 4.8, GPT-5.6 Sol, and GPT-5.5 across nine benchmarks in the categories of coding agents, general agents, and visual agents.
In Meituan's own evaluation, LongCat-2.5 improves substantially over its predecessor and comes close to Claude Opus 5 and GPT-5.6 Sol on several agent benchmarks. | Image: Meituan

The harder variant, Terminal Bench v4.0, shows the distance more clearly. LongCat-2.5 scores 38.5, slightly ahead of GPT-5.6 Sol's 37.3 but well behind Claude Opus 5's 53.9. On BrowseComp, a web search benchmark, it posts 91.9, the highest number in the table. Meituan ran that test in its own environment, so the score can't be compared directly with results from standard harnesses.

Access, licensing and cost

The weights are available on Hugging Face under the MIT license, which permits commercial use, modification and redistribution. People who want to test the model first can use the chat interface at longcat.ai. The API supports both the OpenAI and Anthropic request formats and works with agent harnesses such as Claude Code, Codex and Hermes, which lowers the cost of trying it inside an existing workflow. At launch, Meituan is charging a limited-time promotional price of roughly €0.27 per million input tokens and €1.07 per million output tokens.

The predecessor, LongCat-2.0, was trained in late June entirely on a cluster of more than 50,000 Chinese AI chips, according to Meituan. Before its release, it circulated anonymously as "Owl Alpha" on OpenRouter, which gave developers a chance to test it without knowing who had built it. Chinese open-weight models such as Kimi K3.1 and GLM-5.3 have also closed much of the gap with top US models on common benchmarks, so LongCat-2.5 is one of several releases pushing in the same direction.

What to watch next

My read is that the hardware story matters more than any single benchmark. Open weights designed for Chinese accelerators from the start make that hardware a workable option for teams outside China, and the 4-bit path with its published throughput numbers gives them something to test. The unanswered questions are about independent evaluation. Meituan's numbers come from its own harness, and the most eye-catching result, BrowseComp, comes from its own environment. Third-party runs on Terminal Bench and DeepSWE, ideally with the latest Claude and GPT models included, will show whether LongCat-2.5 is close to the frontier or only close on the company's chosen tests.

Meituan is also working on image generation. It previously released LongCat-Image, an open-source text-to-image model with 6 billion parameters that prioritizes efficiency over raw size. If the company brings the same efficiency focus to its multimodal agent models, the open-weight Ascend-ready stack could become a practical option for teams that have been waiting for a credible alternative to Nvidia-based serving.

Source: Jonathan Kemper · the-decoder.com
Arrived · Hpjadul

Also departing from this gate