On November 2, Meituan released LongCat-2.5, its new flagship large model. It is a natively multimodal model built for long-horizon agent tasks, with 1.6 trillion total parameters, and it was trained end to end on domestic computing power.

The previous generation, LongCat-2.0, arrived in June 2026 and was the first trillion-parameter model to complete both training and inference entirely on domestic computing power. Version 2.5 pushes that line one step further: the goal is no longer just to prove a model can be trained on domestic chips, but to fit multimodality and long context in as well.

1.6 Trillion Parameters, Landing Between Two Claudes

Meituan says LongCat-2.5 not only supports multimodal input natively but also matches the mature international closed-source models and leading open-source models currently on the market.

The benchmark framing is specific: across a range of mainstream benchmarks, LongCat-2.5 scores between Claude-Opus-4.8 and Claude-Opus-5, and surpasses Claude-Opus-5 on some of them.

The positioning is worth noting. Meituan is not claiming across-the-board superiority; it places the model precisely between two generations of a closed-source flagship, clearing the newer one only on selected tests. That is unusual restraint for a model vendor.

The model is already available to try at LongCat.ai, and developers can access it through an API that launched alongside it.

The Training Bottleneck Was Per-Chip Memory

Most of the engineering substance in this model sits in the training stage.

During mid-training, LongCat-2.5 ran long-context training on sequences of up to 1M tokens, for a cumulative total in the trillions of tokens. Mid-training sits between pre-training and post-training, and typically involves running the model through another round on higher-quality, longer data.

The difficulty is that when multiple accelerators work together on sequences that long, memory capacity and chip-to-chip communication become the immediate constraint on efficiency. The practical limitation of domestic computing power is smaller per-chip memory: if the model does not fit, it has to fetch data across chips constantly, and once communication overhead climbs, the compute is wasted.

Meituan's answer was to restructure task allocation around the superpod, keeping most of the communication for parameters and task data inside the superpod's high-speed interconnect rather than sending it back and forth across the wider cluster. In plain terms, the chips that need to talk to each other most often are placed on the same high-speed path.

The result, as stated in the company's technical blog: on the long-context supervised fine-tuning tasks tested, per-chip training throughput rose 86% over a DDP baseline. On a per-chip throughput basis, that means one chip now does roughly the work of 1.86 chips in the same amount of time. DDP, or distributed data parallel, is the most common form of distributed data parallelism, and using it as the reference point indicates the gains really did come from the communication and memory layer.

Inference Cut to 4 Bits, Throughput Up Another 40%

On the inference side, LongCat-2.5 uses a W4A4 scheme, quantizing both expert weights and activations to 4 bits.

Quantization represents the numbers used in computation at lower bit widths. Fewer bits mean cheaper data movement and computation, but accuracy becomes easier to lose. The common industry approach is to compress weights aggressively and keep activations more conservative, such as W4A8. Taking both down to 4 bits is considerably harder.

According to Meituan, the scheme optimizes computation and data transfer around the hardware characteristics of domestic AI accelerators, and on the downstream tasks evaluated it holds performance on par with the unquantized model.

The production stress-test numbers: compared with a W8A8 INT8 baseline, where both weights and activations are 8 bits, throughput rose roughly 40% for both input processing and per-token generation; time to first token fell about 16%, and per-token generation time thereafter fell about 20%. Time to first token and per-token latency are the two metrics users feel directly — one determines how long before text starts appearing, the other how fast it comes out.

Not a Token Factory, but a Tool for Merchants

Wang Xing laid out the thinking behind this model on the company's second-quarter earnings call on August 28, 2026.

His framing then: the long-term value of AI capability lies in whether it can genuinely enter real workflows, solve real problems, and build into durable infrastructure. "We will not become a token factory. Our models and AI products will be used to support our core business, improving the experience for users and merchants and the company's operating efficiency."

In service of that core business, LongCat-2.5 was tuned heavily for business analysis and software development. The technical blog offers one example: in a multi-channel brand performance analysis, the model independently built a channel profit statement, identified a growth-quality problem where payment GMV rose 23.2% while channel profit fell 1.7%, then factored in clearance economics, repeat purchase behavior and inventory risk to propose a differentiated resource allocation plan — carrying the task from accounting through diagnosis to recommendation.

On the product side, Meituan has launched Miaoshou, an AI operations workbench built on the LongCat series. It helps merchants interpret industry metrics and brand rules in high-value operating scenarios, and, within the permissions merchants grant, calls tools and systems to carry out operational tasks. Miaoshou is already in use at chains including Guming, Xushengji and Yihetang, covering brand performance analysis, store management and customer acquisition.

Put the two threads together and LongCat-2.5's positioning comes into focus: on one side, pushing model capability close to the international front rank using domestic computing power; on the other, embedding that capability into the local-services operating workflows Meituan knows best, rather than selling tokens as a general-purpose business.