Meituan ships LongCat-2.5, a 1.6T-param multimodal agent model
TL;DR
- Meituan trained LongCat-2.5 entirely on domestic Chinese AI accelerators, producing a frontier-scale MoE model without Nvidia hardware.
- The training pipeline converts real agent breakage logs into curriculum tasks, targeting agentic reliability rather than raw benchmark scores.
- Meituan claims LongCat-2.5 beats GPT-5.6-Sol on 20 of 27 shared tests and Claude Opus 5 on 10, with no independent evaluation yet published.
Meituan on November 2 released LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts model aimed at long-horizon agent work, with roughly 48 billion parameters active per token. TechNode's writeup describes the model as "designed for tasks requiring sustained execution by AI agents."
Other outlets covering the preview add one number TechNode leaves out: a 1-million-token context window Meituan is pitching for agents that have to hold terminals, browsers, GUIs, spreadsheets and design tools in working memory across many steps. Crypto Briefing also reports a 128,000-token maximum output and W4A4 quantization, with training on domestically produced Chinese AI accelerators.
The training notes are where this gets interesting. Meituan says it "expanded training on long task sequences" and "converts failures encountered by real agents into reproducible training tasks," feeding its own agent breakage back in as curriculum. Vision is folded in at the token level rather than bolted on: the model "processes visual and text inputs within a unified token framework," supporting what the company calls visual coding, creative design, and 3D tasks.
For now LongCat-2.5 is live through Meituan's online chat service and API, with GitHub and Hugging Face releases planned to follow. It is our third China-AI story today and part of a steady run of 289 pieces on that beat over the last 90 days, which is the real backdrop here more than any single benchmark number.
What others are reporting
-
Crypto Briefing Read →
Leads with domestic-chip training as the primary story, explicitly framing LongCat-2.5 within US export restriction context rather than as a product launch.
LongCat-2.5 was trained on domestically produced Chinese AI accelerators.
-
RuntimeWire Read →
Focuses on commercial infrastructure: cached-input pricing and 1M-token window as developer-adoption levers, framing capability claims through cost and deployment practicality.
LongCat-2.5 shows a large platform company pushing frontier-scale agentic coding with native vision, a 1M-token window, and low cached-input pricing.
-
Emergent Read →
Details the agent-failure curriculum: real breakage logs fed back as training tasks, plus native token-level vision integration rather than a bolted-on encoder.
The model converts failures encountered by real agents into reproducible training tasks, feeding breakage logs back into curriculum.
Originally reported by technode.com
Read the original article →Original headline: Meituan Opens LongCat-2.5 Preview: 1.6T-Param Native Multimodal Agentic Model With 1M-Token Context