LongCat-2.5's Weights Are Now Open: A 1.6T-Parameter Agent Model, Trained Without a Single Nvidia GPU

AI disclosure: This article was drafted with AI assistance and checked against The Decoder's report on the weights release, published 15 November 2026. I have not run LongCat-2.5 hands-on — the weights only just landed, and self-hosting one takes hardware almost nobody has at home. Nothing here is sponsored, and there are no affiliate links on this page.

Meituan released the open weights for LongCat-2.5 today, November 15, 2026 — and the headline isn't just the model. It's how it was built: a 1.6-trillion-parameter model for autonomous AI agents, trained entirely on Chinese hardware, with zero Nvidia GPUs anywhere in the process.

This is the follow-up to the announcement from a week ago, on November 2, when LongCat-2.5 was unveiled as a model built for long, multi-step agent work — especially coding. Today's news is the part that changes who can actually use it: the weights are downloadable now, under the MIT license, which means commercial use is allowed. But "downloadable" and "usable" are two very different things here, and the gap between them is the real story.

What was released today

The weights release covers the full LongCat-2.5 model: a mixture-of-experts design with roughly 1.6 trillion parameters in total, of which about 48 billion are active per token. That architecture is the reason a model this size is practical at all — most of the parameters sit idle for any given token, so inference costs stay closer to a 48B model than a 1.6T one.

It also adds image understanding on top of LongCat-2.0, converting images into tokens so the model can work with visual input alongside text. For an agent model, that matters: agents that can see a screenshot, a diagram or a UI mockup can take on tasks a text-only model can't even start.

The license is the MIT license — the permissive kind. You can use the weights commercially, build products on them, fine-tune them. There is no "research only" catch hiding in the fine print.

The benchmark claim — and the grain of salt

Meituan reports a Terminal-Bench score of 88.4 for LongCat-2.5, which would put it just behind Claude Opus 5 on that benchmark. Terminal-Bench is a reasonable yardstick for agent work: it tests whether a model can actually complete multi-step tasks in a terminal, not just answer questions about them.

But that number is Meituan's own claim, measured by Meituan, on a benchmark they chose to highlight. Nobody independent has had the weights long enough to reproduce it yet, and Terminal-Bench is one benchmark — a strong score there says the model is good at terminal-based agent tasks, not that it's the best model at everything. Treat 88.4 as a promising signal, not a settled fact, until independent evaluations land.

The no-Nvidia part is the real headline

Here's the detail that makes this release genuinely unusual: LongCat-2.5 was trained entirely on Chinese hardware. Its predecessor, LongCat-2.0, was trained on more than 50,000 Chinese AI chips. The sample deployment Meituan describes runs on 194 Huawei Ascend Atlas A2 accelerators.

And there is, apparently, no guidance at all for running it on Nvidia GPUs.

That has two consequences, and they point in opposite directions. On one side, it's a proof point: a frontier-class agent model can now be trained without the hardware everyone assumed was indispensable. That changes the strategic picture for anyone watching the AI hardware race. On the other side, it means the weights release is far less useful to most developers than it looks at first glance. The overwhelming majority of self-hosting setups — and most cloud GPU capacity — runs on Nvidia. A model with no Nvidia path is, for most people, a model you can read about but not run.

What it actually takes to self-host

Let's be blunt about this, because "open weights" gets people excited in a way the practical details don't support:

• This is not a laptop model. 1.6 trillion parameters, even with only 48B active per token, needs serious accelerator hardware and a lot of memory. The sample deployment Meituan cites uses 194 Huawei Ascend Atlas A2 accelerators. That's a cluster, not a workstation.
• You need Ascend hardware, or the patience to port it. With no Nvidia guidance from Meituan, getting this running on the GPUs most developers actually have is a community project that hasn't happened yet. Someone may well do it — open weights invite exactly that kind of work — but today, on release day, it's not a given.
• The realistic route for most developers is the API. Meituan's launch promo pricing is roughly €0.27 per million input tokens and €1.07 per million output tokens. That's cheap for a model of this class, and it's the way almost everyone who tries LongCat-2.5 in the next few weeks will actually use it.

None of that makes the release unimportant. Open weights under a permissive license, for a model that scores near the top of an agent benchmark, is a real event — it just matters most to labs, cloud providers and teams with Ascend access right now, and to everyone else a little later, once the community figures out the porting.

What it means for developers building agents

If you're building agent products, the practical takeaway is threefold:

1. The open agent-model field just got a serious new entrant. Until now, the strongest open-weights agent options were a clear step behind the closed frontier models. If the Terminal-Bench number holds up independently, that gap just narrowed a lot — and an MIT license means you can build a business on it.
2. Pricing pressure is real. At roughly €0.27 per million input tokens, hosted LongCat-2.5 undercuts most frontier agent APIs by a wide margin. Even if you never self-host, cheap capable agent inference changes what kinds of always-on agent products are economically viable.
3. Hardware diversity is now a real variable in your planning. A frontier model that only runs well on Ascend hardware is a bet on a second hardware ecosystem. For most teams that's not a reason to switch anything today — but it's a reason to stop assuming Nvidia is the only path to frontier performance.

The honest limits

• No hands-on evidence yet. Everything above comes from The Decoder's report on the release, published today, 15 November 2026. I haven't run the model, and neither has anyone writing about it on day one.
• The benchmark is the vendor's own. Terminal-Bench 88.4 is Meituan's number. Independent reproduction is the test that matters, and it hasn't happened yet.
• Self-hosting is a specialist job right now. No Nvidia guidance, a 194-accelerator sample deployment, and a model this size mean "open weights" does not mean "runs on your machine."
• The API is the practical route, and promo pricing is promo pricing. Launch pricing at €0.27/€1.07 per million tokens may not be where the price settles. Check the current rate before you build a cost model on it.

My take

There are two stories here, and they're both true at the same time. The first is that a 1.6T-parameter agent model with a near-frontier benchmark score is now downloadable under the MIT license, and that's a genuine milestone for open AI — the kind of release that would have been unthinkable as a fully open, commercially usable model two years ago.

The second is that "open" has a hardware asterisk the size of a data centre. A model trained without Nvidia, with no Nvidia deployment path, is open in the legal sense and gated in the practical sense for almost everyone who might want to run it. The community will probably close that gap — open weights have a way of getting ported — but on release day, the honest description is: free to use, expensive to run, and cheapest through Meituan's own API.

The thing to watch over the next two weeks isn't the announcement cycle. It's whether independent evaluators reproduce that 88.4, and whether someone gets LongCat-2.5 running well on hardware the rest of us can actually rent. If both happen, this stops being a China-hardware story and becomes an everyone story. If neither happens, it's still an important proof point — just one most developers will experience through an API, not on their own machines.

Have you tried it yet — and if the weights were ported to your hardware tomorrow, would you self-host it or just call the API?

Sources: The Decoder — "Meituan releases a 1.6 trillion parameter agent model that never touched an Nvidia GPU" (https://the-decoder.com/meituan-releases-a-1-6-trillion-parameter-agent-model-that-never-touched-an-nvidia-gpu/), published 15 November 2026. All capability, benchmark, hardware, licensing and pricing claims are Meituan's as reported there, stated as such. No hands-on testing was performed for this article. This breaking article is not the daily buyer-intent post, and it contains no affiliate links.

FAQ

What is LongCat-2.5?
LongCat-2.5 is Meituan's 1.6-trillion-parameter mixture-of-experts model built for autonomous AI agents — long, multi-step tasks, especially coding — with about 48 billion parameters active per token. The model itself was announced on 2 November 2026; on 15 November 2026 Meituan released its open weights under the MIT license.

What does the MIT license mean for LongCat-2.5?
The weights are released under the MIT license, a permissive open-source license that allows commercial use — you can build products on the model and fine-tune it without a "research only" restriction.

Can I run LongCat-2.5 on Nvidia GPUs?
Not straightforwardly, at least not yet. LongCat-2.5 was trained entirely on Chinese hardware with zero Nvidia GPUs — LongCat-2.0 was trained on 50,000+ Chinese AI chips — and Meituan gives no guidance for Nvidia GPUs. The sample deployment runs on 194 Huawei Ascend Atlas A2 accelerators. Community ports may appear, but on release day there is no official Nvidia path.

How good is LongCat-2.5?
Meituan reports a Terminal-Bench score of 88.4, just behind Claude Opus 5. That is Meituan's own claim and has not yet been independently reproduced, so treat it as a promising signal rather than a settled fact.

How much does the LongCat-2.5 API cost?
Launch promo pricing is roughly €0.27 per million input tokens and €1.07 per million output tokens. Promo pricing may change, so check the current rate before building on it.

How is the weights release different from the 2 November announcement?
On 2 November 2026 Meituan announced the LongCat-2.5 model itself. On 15 November 2026 it released the actual model weights for download under the MIT license — the step that lets others run, fine-tune and build on the model rather than only reading about it.