Meituan has released LongCat-2.5, a new large language model built specifically for long-running agent tasks. According to the Chinese company, the model has 1.6 trillion parameters, understands images natively and was trained entirely on Chinese AI accelerators. Developers can use it right away through an OpenAI-compatible API at a limited-time launch price of $0.30 per million input tokens.
In the benchmarks Meituan published, LongCat-2.5 mostly lands between Claude Opus 4.8 and Claude Opus 5, and beats Opus 5 on a few tests. Unlike its predecessor LongCat-2.0, there are no open weights this time. Here is what is behind the model and what it means for developers.
LongCat-2.5: the key facts
According to the LongCat team’s technical blog post dated November 1, LongCat-2.5 is a mixture-of-experts (MoE) model with 1.6 trillion total parameters. Around 48 billion of them are active per token. The model builds on LongCat-2.0 and is meant to improve in two areas above all: understanding images and writing code autonomously over long stretches.
Meituan’s announcement from November 2 says training ran entirely on domestic Chinese compute. Meituan already describes LongCat-2.0 as the first trillion-parameter model it trained and served end to end on Chinese chips, and calls the approach “domestic models on domestic chips”. The company does not say which accelerators it used. The blog only mentions cards with less than 80 GB of memory and so-called supernodes, tightly interconnected groups of accelerators.
During mid-training, the team used sequences of up to one million tokens. The example response in the API documentation lists a context length of 1,048,576 tokens.
How LongCat-2.5 handles images
The way the model processes images is technically interesting. LongCat-2.5 turns visual signals into multi-level discrete tokens, much like text is split into word tokens. The team applies residual vector quantization to features extracted at native resolution. The model’s backbone then handles text and images in the same token format.
That has practical benefits for training. Because the image tokenizer is fixed, image tokens can be precomputed and cached offline. The vision encoder drops out of the training graph, and the load across compute nodes becomes more even. Meituan says the main beneficiaries are tasks where the model turns visual input into code, such as interactive web pages, 3D showcases or design drafts.
Training for long-horizon agent tasks
Most of the blog post covers reinforcement learning for long tasks. Part of the training data is generated automatically: a pipeline filters GitHub pull requests and linked issues, lets agents build tasks and test environments from them, and validates each one with tests that must fail before the fix and pass after it. Internal Meituan coding tasks are added on top. Failed agent runs from real use are turned into new, verifiable training tasks.
The team also took steps against reward hacking, where a model games the scoring. Training environments have no network access, and git history and cached files are removed. A separate language model audits runs for exploits. Extra penalties discourage repetitive reasoning and duplicate tool calls.
Benchmarks: between Opus 4.8 and Opus 5
The numbers come from Meituan itself. Scores for other models were partly taken from public reports and partly measured by the team, which explicitly calls its own measurements indicative only. Some examples from the table:
- Terminal Bench v2.1: LongCat-2.5 scores 88.4. Claude Opus 5 is at 89.1, GPT-5.6-Sol at 88.8 and DeepSeek-V4.1-Flash at 90.6.
- Terminal Bench v4.0: 38.5 for LongCat-2.5, compared with 53.9 for Opus 5, 41.8 for GLM-5.3 and 37.3 for GPT-5.6-Sol. Its predecessor LongCat-2.0 managed only 3.0.
- DeepSWE v1.1: 64.6, versus 68.8 for Opus 5 and 72.7 for GPT-5.6-Sol.
- BrowseComp: 91.9, narrowly ahead of Opus 5 at 90.8.
- MCP Atlas: 83.1, versus 85.8 for Opus 5.
According to the footnotes, most coding tests ran in the Claude Code harness. The jump over LongCat-2.0 on agent tasks is large. Still, the model is not at the top: on the harder coding benchmarks, Claude Opus 5 leads clearly. Newer models such as GPT-6.1 are not in the comparison table, and there are no independent measurements yet.
Faster inference with 4-bit quantization
For serving, Meituan relies on W4A4 quantization: expert weights and activations are reduced to 4 bits and computed on INT4 units. The team says the model matches its unquantized version on all evaluated downstream tasks. Compared with a W8A8 INT8 baseline, online stress tests showed 43.3 percent higher prefill throughput and 42.77 percent higher decode throughput. Time to first token fell by 15.91 percent and time per output token by 19.72 percent.
The team also reports training gains. A customized FSDP setup reached 86 percent higher per-device throughput than a DDP baseline on long-sequence fine-tuning. That is worth watching for anyone following how Chinese labs work around US export restrictions on AI chips.
Access and pricing for developers
LongCat-2.5 can be tried in the chat at longcat.ai and called through the API. The API documentation describes an OpenAI-compatible chat completions endpoint with streaming and a parameter that turns reasoning on or off. There is also an Anthropic-compatible endpoint, so existing tools can often be connected with little work. With thinking enabled, temperature is fixed at 1.0.
The pricing page lists limited-time discounted prices per million tokens:
- Uncached input: $0.30 (¥2)
- Cached input: $0.006 (¥0.04)
- Output: $1.20 (¥8)
The page does not say how long the discount lasts or what comes after it. For comparison, Kimi K3.1 Pro and DeepSeek V4.1-Flash also compete on aggressive pricing. Competition among Chinese providers is pushing down the cost of agentic work noticeably.
One caveat: the example response of the model endpoint currently lists text as the only input modality. The documentation does not make clear whether images can already be sent through the API. Teams handling personal or confidential data should also review the provider’s data protection terms before using the API in production.
No open weights
Unlike LongCat-2.0, whose weights are on Hugging Face, Meituan has not released weights for LongCat-2.5 so far. On November 2, the team’s Hugging Face account did not list such a model. Whether an open version will follow is unclear. The demo videos announced in the blog post are not online yet either.
Our take
For Meituan, LongCat is not an end in itself. CEO Wang Xing said in August that the company does not want to become a “token factory” and that its models should support the core business. LongCat already powers “Miaoshou”, an AI workbench that helps merchants with business analysis and store management. The API is more of a by-product, if a cheap one.
For developers, LongCat-2.5 is mainly another sign of how quickly Chinese labs are catching up on agent models, and on their own hardware. If you are looking for low-cost models for long coding or research tasks, the compatible API makes LongCat-2.5 easy to try. Treat the benchmarks as vendor claims and check them against your own tasks. How capable open and cheap models have become was also shown recently by the report on GLM 5.3 and the patch gap.
