Independent/Reader-funded/Infrastructure, not tokens
DeAINEWS

AI you control — open models, private inference, and the networks that run them.

Open-Weights Releases

Meituan ships LongCat-2.5: 1.6T-param open-weights agent model, MIT

Meituan released LongCat-2.5 open weights under MIT: a 1.6T-parameter multimodal MoE with about 48B active per token, reportedly trained on Huawei Ascend chips.

DeAI is powered by Morpheus (mor.org). We cover competing providers on the same terms — see our methodology.

A datacenter rack aisle with accelerator nodes and amber status lights, illustrating Meituan's LongCat-2.5 open-weights model reportedly trained on Huawei Ascend hardware. Illustration: DeAI
A datacenter rack aisle with accelerator nodes and amber status lights, illustrating Meituan's LongCat-2.5 open-weights model reportedly trained on Huawei Ascend hardware. Illustration: DeAI

Meituan's LongCat team has released the weights of LongCat-2.5, a 1.6-trillion-parameter multimodal mixture-of-experts model built for long-horizon agent work, under the MIT license on Hugging Face. It matters because frontier-scale agent weights you can self-host are now priced at commodity-API levels, and Meituan says the whole training run stayed off Nvidia hardware.

Key facts

What happened

The LongCat tech blog describes LongCat-2.5 as a native multimodal model: images are converted into multi-level discrete tokens and processed in the same token stream as text, rather than through a bolted-on vision encoder. The team says this design targets visual coding, design, and 3D tasks alongside long-horizon agentic coding.

Training leaned on large-scale agentic reinforcement learning. The team synthesized tasks from GitHub pull requests and real internal coding work, built sandboxed environments with network access disabled and git history removed, and used fail-to-pass and pass-to-pass tests plus an LLM judge to catch reward hacking. Specialized teacher models for agent behavior, reasoning, and instruction following were distilled into the main model through full-vocabulary KL divergence.

Meituan quantized the expert weights and activations to 4-bit (W4A4) so Chinese accelerators can run the model natively. The company reports 43.3 percent higher prefill throughput and 42.77 percent higher decode throughput versus its W8A8 baseline, with accuracy it describes as on par. The W4A4 build ships in a separate repository, and the self-hosting guide covers only Ascend hardware, with no Nvidia instructions published.

On benchmarks, the numbers in Meituan's own table are self-reported and marked as indicative. LongCat-2.5 scores 88.4 on Terminal Bench v2.1, close to Claude Opus 5 at 89.1 and GPT-5.6 Sol at 88.8, while DeepSeek-V4.1-Flash leads that table at 90.6. On the harder Terminal Bench v4.0 the gap is wider: 38.5 for LongCat-2.5 against 53.9 for Opus 5. DeepSWE improved from 9.7 in LongCat-2.0 to 64.6, which still trails every other model in the table except Claude Opus 4.8. The standout is BrowseComp at 91.9, the highest score in the table, though Meituan ran that test in its own harness.

Why it matters

For builders, the practical fact is not the benchmark table. It is that a model at this scale and capability tier ships under MIT, the most permissive mainstream license available, so there are no usage caps or country restrictions to lawyer through before self-hosting. That is a different proposition from the capped community licenses some frontier-scale open-weight releases still carry, and it puts LongCat-2.5 alongside releases like Mistral Large 4 in the fully-permissive tier.

The second fact is the hardware story. Meituan states that training and inference optimization ran on domestic Chinese accelerators, and the deployment guide exists only for Ascend. THE DECODER frames the release as a model that never touched an Nvidia GPU; that framing rests on Meituan's own statements, which we have not independently verified. If accurate, it is further evidence that open-weight releases optimized for Chinese silicon from the start are becoming a distinct engineering track, not a porting afterthought. Chinese open-weights releases such as Kimi K3.1 Pro have already closed much of the benchmark gap with US frontier models.

The third fact is price. A promotional API rate near €0.27/€1.07 per million tokens sits at the aggressive end of the commodity tier, and the OpenAI-compatible plus Anthropic-compatible endpoints mean most existing harnesses can point at it without code changes. Cheap frontier-adjacent capacity with permissive weights widens real options for multi-provider setups and self-hosting evaluation alike.

Background

LongCat-2.5 builds on LongCat-2.0, which Meituan says it trained in late June on a cluster of more than 50,000 Chinese AI chips, and which briefly ran anonymously on OpenRouter under the alias Owl Alpha. The 2.0-to-2.5 jump is largest on agent benchmarks: Terminal Bench v2.1 moved from 70.8 to 88.4, and DeepSWE from 9.7 to 64.6, according to Meituan's evaluation table.

The release fits the pattern we track in what open-weight models are and why they matter: frontier-lab capability arriving as downloadable weights, with the license file carrying as much builder significance as the benchmark table. It also follows the same-week release rhythm of other Chinese labs; the weights landed on Hugging Face roughly a week after the model's unveiling, and the repo counted 1,573 downloads at the time of writing.

One scope caveat deserves plain language. The blog and model card describe W4A4 inference and NPU deployment paths in detail, while guidance for running the BF16 weights on Nvidia CUDA or AMD ROCm is absent so far. Community ports typically follow for MIT-licensed models of this size, but nothing is published yet, so teams without Ascend hardware should budget for integration work before committing.

What's next

Watch three things. First, independent benchmark reproduction: Terminal Bench v2.1 at 88.4 and BrowseComp at 91.9 are self-reported until a third party reruns them, and the closest comparison points (GPT-6.1 Sol, Claude Opus 5.5) do not appear in Meituan's table at all. Second, a CUDA or ROCm serving path, which would tell you how much of the Ascend-only posture is engineering choice versus constraint. Third, the post-promo API price: the launch rate is explicitly limited-time, and no successor rate has been published.

Questions

Is LongCat-2.5 really open weights?
Yes. The Hugging Face repo meituan-longcat/LongCat-2.5 ships the weights under the MIT License, which permits commercial use, modification, and redistribution. A separate repo carries the W4A4-quantized variant for NPUs.
Was LongCat-2.5 trained without Nvidia GPUs?
Meituan says training and long-context inference optimization ran on domestic AI accelerators with under 80 GB of memory per device. No independent audit of the training cluster exists, so treat the Ascend-only claim as a vendor report, not a verified fact.
How much does the LongCat-2.5 API cost?
At launch Meituan charges a limited-time promotional price equivalent to roughly €0.27 per million input tokens and €1.07 per million output tokens, according to THE DECODER. No standard post-promo rate has been published.
How good is LongCat-2.5 on coding benchmarks?
Meituan reports 88.4 on Terminal Bench v2.1 and 91.9 on BrowseComp. Those are self-reported numbers from Meituan's own evaluation table; independent reproductions have not been published yet.

Sources

  1. Introducing LongCat-2.5 (tech blog) — LongCat (Meituan)
  2. meituan-longcat/LongCat-2.5 model card — Hugging Face
  3. Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU — THE DECODER
  4. LongCat API Platform Quick Start — LongCat (Meituan)

About DeAI

DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.

Powered by Morpheus and StrandCMS

Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more about the Morpheus Inference API →

Sponsor disclosure — not editorial

Powered by Morpheus and StrandCMS. Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more →