A footprint of 1.6 trillion parameters now runs entirely decoupled from Western accelerator supply chains, following the public release of the weights for LongCat-2.5 by Chinese food delivery company Meituan. Built specifically for autonomous AI agents that handle multi-step workflows, the system challenges the operational assumption that frontier-scale agentic models require Western accelerators to train and deploy. The release demonstrates how enterprise architecture can adapt when physical hardware constraints collide with high compute demands.
Whether an enterprise can deploy a competitive frontier architecture without standard accelerator clusters comes down to physical execution and unit cost. Meituan has structured the system to prove that non-Western hardware clusters can sustain extreme parameter volumes without collapsing under inference overhead. This operational pivot alters the cost calculations for enterprise builders evaluating self-hosted models against proprietary infrastructure.
Sparse Routing and Cluster Topologies
The architectural core of LongCat-2.5 relies on a mixture-of-experts design that activates roughly 48 billion parameters per token out of its 1.6 trillion total parameters. Rather than engaging the entire network for every computation, the system routes input through specialized subnetworks. As the documentation for Meituan LongCat-Flash shows, the system scales total parameters significantly while maintaining a sparse runtime profile. Through the API, the system handles context windows of up to one million tokens at once and generates continuous responses up to 128,000 tokens long.
Instead of treating multimodal tasks as an operational afterthought, the model integrates visual data directly into its text stream. LongCat-2.5 converts images into tokens that the model processes alongside text, avoiding secondary modular pipelines. The hardware footprint required to run this engine bypasses Western suppliers entirely.
The self-hosting guide describes a cluster of 194 Huawei Ascend accelerators from the Atlas A2 series.
In that published documentation, there is no guidance for Nvidia GPUs. The deployment allocates two processors to image handling, 64 to prompt ingestion, and 128 to response generation. To fit within this hardware profile, Meituan quantizes the experts' weights and activations to 4-bit precision, an execution format processed natively by the silicon. According to Meituan's load tests, this compression increases prompt processing throughput by 43.3 percent and response generation throughput by 42.77 percent compared with an 8-bit baseline, while keeping measured accuracy on par with the uncompressed model.
Benchmark Limits and Infrastructure Costs
The fundamental operational counter-argument to domestic hardware training centers on execution quality in complex software workflows. The LongCat team reports 88.4 points on Terminal Bench v2.1, a benchmark that measures how well agents solve command-line tasks, placing the system behind Claude Opus 5 at 89.1, GPT-5.6 Sol at 88.8, and Deepseek-V4.1-Flash at 90.6. On DeepSWE, which tests how models fix bugs in real code repositories, LongCat-2.5 achieves 64.6 points, trailing GPT-5.6 Sol at 72.7 and Deepseek-V4.1-Flash at 74.2 while outperforming Claude Opus 4.8.
Through the API, the model handles up to one million tokens at once and generates responses up to 128,000 tokens long.
On the more demanding Terminal Bench v4.0 evaluation, LongCat-2.5 marks 38.5 points, slightly outpacing GPT-5.6 Sol at 37.3 yet lagging Claude Opus 5 at 53.9. Performance holds closer on web tasks, where the model registers 91.9 on the BrowseComp search benchmark under Meituan's internal testing environment. The data shows that while domestic hardware stacks can close the execution gap on general terminal scripting, complex repository-level debugging exposes persistent software engineering deficits.
Licensing and pricing structures dictate the commercial viability of deploying alternative stacks. Meituan has released the weights under the MIT license, permitting commercial modification and self-hosted redistribution. For external developers utilizing API access, the company set an introductory promotional price equivalent to roughly €0.27 per million input tokens and €1.07 per million output tokens, backed by earlier infrastructure testing where Meituan trained predecessor LongCat-2.0 in late June entirely on a cluster of more than 50,000 Chinese AI chips.
Side by side
In early August the developer's system relied on 69 billion parameters with just 3 billion activated per token, whereas this deployment scales that sparse approach to 1.6 trillion total parameters entirely outside foreign supply chains. That architecture successfully sustains a million-token context window, yet it still trails rivals like GPT-5.6 Sol in real-world software debugging on DeepSWE. Can Meituan bridge the performance gap to match Deepseek-V4.1-Flash's 74.2 score on complex coding benchmarks while remaining confined to domestic silicon?
Our stories: Meituan LongCat-Flash: High-Efficiency Sparse MoE AI (August 1).