A very large no-Nvidia speedrun
Meituan’s LongCat team has pushed LongCat-2.5 into the spotlight by releasing the weights for a 1.6-trillion-parameter agent model that, according to the launch coverage, was trained entirely on Chinese hardware rather than Nvidia GPUs . The key point is not only the headline parameter count. It is the combination of frontier-scale ambition, agent-oriented design, open weights, and a hardware stack that does not start from Nvidia’s CUDA ecosystem.
The release matters because Nvidia accelerators have become the default substrate for training large AI systems. That default has created familiar bottlenecks: constrained supply, high capital costs, export-control pressure, and a software ecosystem that makes switching away from Nvidia difficult. LongCat-2.5 does not end that dependency for the industry. But it is a public proof point that a large Chinese AI team can train and release a trillion-scale agent model on a different stack, then document a self-hosting route that points to Huawei Ascend hardware rather than Nvidia GPUs .
The current state of the story is therefore two-sided. On one side, Meituan has shipped a model with open weights, a permissive MIT license, and ambitious agent benchmarks. On the other, most of the performance numbers remain Meituan’s own, and the reference deployment is far beyond the reach of ordinary developers.
What LongCat-2.5 is
LongCat-2.5 is a mixture-of-experts model. Instead of activating all 1.6 trillion parameters for every token, it activates roughly 48 billion parameters per token, letting the model carry a huge total capacity while keeping per-token compute closer to a much smaller dense model . That design is now common among large frontier-style systems, because it helps scale model capacity without making every inference step maximally expensive.
The model is aimed at long-horizon agentic work, especially coding. The reported use cases include multi-step software tasks, complex codebase development, full-project builds, and workflows where an AI system plans, writes code, runs tools, observes failures, and continues iterating . It also supports multimodal inputs. The Decoder reports that the model can handle text, images, and video as inputs, while responding in text .
A notable architectural detail is how LongCat-2.5 handles images. Instead of treating vision as an afterthought bolted onto a language model, the system converts images at native resolution into discrete tokens that are processed alongside text . That design is meant to support visual coding, design interpretation, and 3D-related tasks, although independent testing will have to determine how robust those abilities are in practice.
The hardware claim is the strategic claim
The most important part of the announcement is the training stack. The Decoder reports that LongCat-2.5 “never touched” an Nvidia GPU and was trained entirely on Chinese hardware . WebPulse’s account adds a useful distinction: Meituan’s own model card language is narrower, referring to domestic accelerator optimization, while the third-party reporting identifies the broader no-Nvidia training claim and points to a Huawei-based self-hosting guide .
That distinction matters. The training claim is about how the model was built. The deployment guide is about how others can run it. In the deployment path reported this week, Huawei appears concretely: the sample cluster uses 194 Ascend Atlas A2 accelerators, with two chips for image processing, 64 for prompt ingestion, and 128 for response generation . Dirty Static also highlights that the guide provides no Nvidia GPU path and that the documented self-hosting setup is a data-center-scale project rather than a casual installation .
Meituan also applies 4-bit quantization to the experts’ weights and activations. The point is to reduce memory movement and exploit native INT4 processing on domestic AI accelerators . In Meituan’s own load tests, that 4-bit version reportedly improved prompt-processing throughput by 43.3% and response-generation throughput by 42.77% compared with an 8-bit version, while keeping accuracy comparable to the uncompressed model . These are vendor figures, not independent measurements.
Open weights, but not easy hardware
The weights are available under the MIT license, which permits commercial use, modification, and redistribution . That makes LongCat-2.5 more open legally than many “available” AI models that come with research-only or field-of-use restrictions. Tokenstrom’s summary emphasizes the same practical tension: the model is open in licensing terms, but not easy for most developers to run because of its size and hardware requirements .
That tension is central to the story. “Open weights” does not mean “runs on your workstation.” A 1.6-trillion-parameter model, even with only about 48 billion active parameters per token, still carries a massive memory footprint. The reference deployment requiring 194 Huawei accelerators makes clear that realistic self-hosting is currently an institutional exercise . For most teams, the practical first contact with LongCat-2.5 will be through an API or through a provider already operating the right hardware.
The API angle is also part of Meituan’s adoption strategy. The Decoder reports that the LongCat-2.5 API supports both OpenAI and Anthropic formats, making it easier to plug into existing agent harnesses such as Claude Code, Codex, and Hermes . That compatibility could reduce software friction even if the underlying hardware remains unfamiliar to many non-Chinese developers.
Benchmarks: impressive, but still vendor-led
On Meituan’s reported benchmark table, LongCat-2.5 appears competitive on several agentic coding tasks. The company reports 88.4 on Terminal Bench v2.1, close to Claude Opus 5 at 89.1 and GPT-5.6 Sol at 88.8 in the same table . On the harder Terminal Bench v4.0, the model reportedly scores 38.5, above GPT-5.6 Sol’s 37.3 but well below Opus 5’s 53.9 . On BrowseComp, Meituan reports 91.9, the highest figure in its table, though The Decoder notes that Meituan ran the test in its own environment .
The DeepSWE result is more mixed. Meituan reports LongCat-2.5 at 64.6, a dramatic improvement over the earlier LongCat score of 9.7, but still below several rival models in the same table . WebPulse similarly frames the benchmarks as promising but not conclusive, noting that unmarked scores in Meituan’s tables are internal and “indicative only” .
That caveat should shape every interpretation of the release. The hardware achievement and the model release are concrete. The performance comparison is still provisional. Independent evaluators need to run the weights, reproduce the harnesses, and test the model on real workloads outside Meituan’s own environment.
Why this launch matters beyond Meituan
LongCat-2.5 arrives at the intersection of three pressures: the race for capable AI agents, the economics of inference, and the geopolitics of chips. If a model of this scale can be trained and deployed on domestic Chinese accelerators, then Nvidia’s role as the default platform becomes less absolute. That does not mean Nvidia is displaced. It means there is now a visible alternative stack tied to an open-weight, agent-focused model.
The release also changes the burden of proof. Before, the claim that Chinese domestic accelerators could support near-frontier training was often discussed as an ambition or a policy objective. With LongCat-2.5, Meituan is presenting a shipped model, downloadable weights, a quantized deployment path, and benchmark claims that outside groups can now inspect .
For buyers and developers, the sensible reading is cautious but serious. LongCat-2.5 is not suddenly an easy replacement for Nvidia-backed models. It is too large for casual self-hosting, its strongest numbers still require independent verification, and the documented hardware path is specialized. But it is also not just a press release. It is a large, permissively licensed agent model built outside the Nvidia training lane. That is why this story matters: Meituan has turned “beyond Nvidia” from a slogan into a testable engineering claim.
Comments
Be the first to comment.