Meituan Just Released a 1.6-Trillion-Parameter Agent Model Trained Entirely on Chinese Chips, and It Scores 88.4 Next to Claude Opus 5 at 89.1
The headline is not the size, it is the hardware
A 1.6-trillion-parameter model is no longer, by itself, a surprise. What makes LongCat-2.5 different is where it was trained and what it was trained to run on.
Meituan, the Chinese food delivery company that also runs the LongCat model program, released the weights for LongCat-2.5 on November 15, one week after the model was officially unveiled. According to Meituan and reporting by The Decoder, the model was trained entirely on Chinese hardware. There is no Nvidia GPU in the story, not in training and not in the sample deployment published alongside the weights. The self-hosting guide describes a cluster of 194 Huawei Ascend accelerators from the Atlas A2 series, and offers no guidance at all for Nvidia hardware.
That is the point of the release. LongCat-2.5 tests whether a frontier-scale agent model can be built, trained, and served end to end on a non-Nvidia stack, close enough to the leading Western models to matter in practice.
What LongCat-2.5 actually is
LongCat-2.5 is a mixture-of-experts model built for autonomous agents, especially agents that work through long, multi-step coding tasks without a human steering every step.
- Total size, active size. The model holds 1.6 trillion parameters in total, but only about 48 billion are active for each token it processes. That split is the whole idea of mixture-of-experts design: keep a large pool of specialized subnetworks, and route each token through only the ones it needs.
- Context and output. Through the API, the model handles up to one million tokens of input at once and can generate responses up to 128,000 tokens long.
- Multimodal input, text output. It understands text, images, and video, and responds in text. Users can toggle its reasoning mode on and off.
- Native image tokens. This version builds on LongCat-2.0 and adds image understanding. Instead of bolting on a separate image encoder after the fact, LongCat-2.5 converts images, at their native resolution, into discrete tokens that the model processes alongside text. Meituan says that approach, drawn from its LongCat-Next research model, is meant to improve visual coding, design, and 3D tasks.
For training, the team used real software tasks drawn from GitHub pull requests, along with recreated agent failures. The detail that matters is how they tried to stop the model from cheating its way through those tasks. In the training environments they blocked internet access, removed Git history and cached files, and used another language model to check solutions for tricks. They also distilled capabilities from several specialized teacher models for agent work, reasoning, and dialog into this one model.
In plain terms, the training setup forced the model to actually solve the problem in front of it, not look up the answer or exploit a shortcut.
The 4-bit choice that makes the Chinese chips work
The hardware story has a technical core, and it is quantization. Meituan quantizes the experts’ weights and activations to 4-bit precision, a format Chinese chips can process natively.
In Meituan’s own load tests, that choice raises throughput by 43.3 percent for prompt processing and by 42.77 percent for response generation, compared with an 8-bit version. Meituan says accuracy stays on par with the uncompressed model. Those are the company’s own measurements, not independently verified, but the logic is clear: if your accelerators are fastest at 4-bit math, you design the model to live at 4-bit, rather than porting a model built for someone else’s hardware and compressing it later.
The sample deployment shows how seriously the company took serving. Of the 194 Huawei Ascend Atlas A2 accelerators in the reference cluster, two handle image processing, 64 handle prompt ingestion, and 128 handle response generation. A separate repository holds the version tuned for this setup.
This is what designing for a hardware stack from the start looks like, as opposed to treating that stack as a compatibility problem to solve later.
How close is it to the Western frontier, on Meituan’s own numbers
All of the benchmark figures below are reported by Meituan itself. They are a reason to test the model, not a final verdict on it, and no independent verification has been published yet.
- Terminal Bench v2.1: 88.4. This benchmark measures how well agents solve command-line tasks. Meituan lists Claude Opus 5 at 89.1, GPT-5.6 Sol at 88.8, and DeepSeek-V4.1-Flash at the top of its table with 90.6. LongCat-2.5 sits just behind Opus 5, close enough that the gap is smaller than the difference between most model generations.
- DeepSWE: 64.6. This test measures fixing bugs in real code repositories. LongCat-2.5 jumps from 9.7 for its predecessor to 64.6 here, a very large generational gain, but it still trails DeepSeek-V4.1-Flash at 74.2 and GPT-5.6 Sol at 72.7 in the same table.
- Terminal Bench v4.0: 38.5. On the harder variant of the terminal benchmark, LongCat-2.5 scores slightly above GPT-5.6 Sol at 37.3, but well below Opus 5 at 53.9. That gap is a useful reminder that a single headline score hides a lot of variation by task difficulty.
- BrowseComp: 91.9. Meituan reports this as the highest score in its table on the web search benchmark, while noting it ran this test in its own environment, a caveat readers should keep attached to the number.
Read together, the pattern is consistent. LongCat-2.5 is not claiming to beat the best Western models across the board. It is claiming, with numbers within a point or two on some agent tests and clearly behind on harder coding tests, that a model trained entirely on Chinese hardware now plays in the same league Claude Opus 5 occupied a few months ago.
The license and the price are part of the strategy
The weights are available under the MIT license, which allows commercial use, modification, and redistribution. Anyone can also try the model in a chat interface at longcat.ai. Its API supports both the OpenAI and Anthropic formats, and Meituan says it plugs into agent harnesses such as Claude Code, Codex, and Hermes.
At launch, Meituan is charging a limited-time promotional price equivalent to roughly 0.27 euros per million input tokens and 1.07 euros per million output tokens, according to The Decoder’s reporting of the launch pricing.
Each choice lowers a different barrier. The MIT license removes the legal question for companies that want to self-host. The dual API formats remove integration work for teams already built on OpenAI or Anthropic interfaces. The promotional price removes the cost excuse for trying it. And the reference deployment on 194 Huawei chips gives a buyer considering Chinese accelerators a worked example, not just a benchmark chart.
Meituan also notes that it trained the predecessor, LongCat-2.0, in late June entirely on a cluster of more than 50,000 Chinese AI chips. That model ran anonymously as Owl Alpha on OpenRouter before it was named. This is not a one-off experiment. It is the second generation of a program testing, in public and at scale, whether the non-Nvidia path holds up.
Why this matters outside China
It is tempting to file this as a China story, and stop there. That would miss two points.
First, the export control logic that has shaped the last three years of AI policy assumes that cutting off access to the best Nvidia chips slows rival model development in a direct, durable way. LongCat-2.5 does not prove that assumption false on its own, and its hardest coding scores still trail the leaders. But a 1.6-trillion-parameter agent model, trained entirely on domestic chips and released openly, is concrete evidence that the constraint pushes work toward alternative stacks rather than simply stopping it.
Second, open-weight models optimized for Chinese chips from the start make that hardware more attractive outside China too. A developer who downloads MIT-licensed weights and finds the best documented deployment path runs on Huawei Ascend hardware now has a commercial reason to look at that hardware. Software ecosystems sell chips as much as chip performance does.
None of this means LongCat-2.5 is the right model for your team today. Its benchmark case rests on Meituan’s own testing, its reference deployment asks for 194 specialized accelerators most companies do not own, and there is no Nvidia deployment guidance for teams that want to run it on hardware they already have.
The bottom line
Meituan released LongCat-2.5 on November 15 with its weights under the MIT license: a 1.6-trillion-parameter mixture-of-experts agent model, about 48 billion parameters active per token, a one-million-token context, native image tokenization, and a 4-bit design built for Huawei Ascend chips, with a 194-chip reference deployment and no Nvidia path documented. On Meituan’s own, not yet independently verified numbers, it scores 88.4 on Terminal Bench v2.1 against 89.1 for Claude Opus 5, and 64.6 on DeepSWE against 74.2 for the table leader. The larger story is the stack underneath it: frontier-scale training and serving, done end to end without Nvidia, now has a public, downloadable proof of concept anyone can inspect and test.