Meituan's LongCat-2.5 Is a 1.6 Trillion-Parameter Agent Model Trained on Chinese Chips; Its Own Charts Put It Behind Claude Opus 5 on the Hardest Coding Test
Meituan released LongCat-2.5 on November 2, a natively multimodal model with 1.6 trillion parameters, about 48 billion active per token, built for long agentic coding tasks. Meituan says it trained and runs the model on domestic AI accelerators with under 80 GB of memory each. Its own benchmark table puts it close to Claude Opus 5 on Terminal Bench v2.1 but well behind on the newer v4.0. It is available by chat and API; weights are planned but not yet out.
The most notable thing about Meituan's new model may be the hardware, not the scores. Meituan released LongCat-2.5 on Monday, November 2, a model with 1.6 trillion total parameters and about 48 billion active per token, aimed at long-running agent work such as building whole software projects. Its technical blog says the model was trained and is served on domestic AI accelerators, each with less than 80 GB of memory. Meituan does not name the chips.
What Meituan says is new#
LongCat-2.5 builds on LongCat-2.0. Meituan says it is natively multimodal: it turns images into discrete tokens so the same backbone handles pictures and text, which it says helps with visual coding, design and 3D tasks. It says it trained the model on more long-horizon tasks with large-scale reinforcement learning, and turns real agent failures into new, verifiable training tasks.
On serving, Meituan says 4-bit weights and activations running on the accelerators' native INT4 units raised prefill throughput by 43.3% and decode throughput by 42.77% against an 8-bit baseline in its stress tests.
How the scores compare, on Meituan's numbers#
Meituan's table compares LongCat-2.5 with OpenAI, Anthropic and Chinese models; scores marked as taken from public reports came from those reports, and the rest Meituan ran itself. On Terminal Bench v2.1, LongCat-2.5 scores 88.4 against 89.1 for Claude Opus 5. On the newer Terminal Bench v4.0, the gap widens: 38.5 against 53.9. On DeepSWE v1.1 it scores 64.6, behind DeepSeek-V4.1-Flash at 74.2, and on BrowseComp it leads the table at 91.9.
Those comparisons are Meituan's alone. Moonshot's Kimi K3.1 Pro, for instance, has an independent ranking, after Artificial Analysis placed it fourth on its Intelligence Index; we found no independent leaderboard result for LongCat-2.5 yet.
Availability#
TechNode reports that the model is available through Meituan's chat service and API platform, with releases on GitHub and Hugging Face planned. The weights were not public when we checked.
Limits#
- All benchmark figures are Meituan's own table; some competitor scores are self-run by Meituan and described by it as "indicative only."
- The blog is dated November 1 and TechNode dates the release November 2. Pricing and API terms were not in either source, and we have not used figures circulating on social media.
- Meituan does not identify the accelerator vendor.