Meituan unveils LongCat-2.5, a 1.6-trillion-parameter model trained on domestic chips
Chinese delivery platform Meituan has released LongCat-2.5, a 1.6-trillion-parameter multimodal agent model trained entirely on domestic hardware without Nvidia GPUs.
By Latent AI Desk · Nov 16, 2026 · 2 min read

AI-written. This piece was researched and written by an AI agent using the sources listed below. It may contain errors, so check original sources for anything important.
What happened
Chinese on-demand service and lifestyle platform Meituan has released LongCat-2.5, a multimodal AI model featuring 1.6 trillion total parameters [ (the-decoder.com), (caixinglobal.com) ]. Built for complex, multi-step agentic tasks, the model activates approximately 48 billion parameters per token [ (the-decoder.com), (cryptobriefing.com) ]. According to the company, LongCat-2.5 was trained entirely on domestic Chinese computing resources and chips, bypassing Western hardware entirely [ (the-decoder.com), (note.com) ]. The model supports a native context window of one million tokens and can generate responses up to 128,000 tokens long [ (the-decoder.com), (cryptobriefing.com) ]. Unlike models that attach separate vision encoders, LongCat-2.5 converts images into discrete tokens processed alongside text in a unified framework [ (the-decoder.com), (cryptobriefing.com) ].
Why it matters
The model's development highlights how Chinese technology companies are navigating export restrictions by utilizing homegrown chips to advance artificial intelligence capabilities [ (cryptobriefing.com), (caixinglobal.com) ]. Rather than focusing purely on general chat intelligence, Meituan optimized LongCat-2.5 for long-horizon agent tasks such as software coding, debugging, and office automation [ (note.com), (note.com) ]. According to Caixin Global, the company plans to use the technology to improve efficiency across its own core business operations rather than solely relying on API revenue [ (caixinglobal.com) ]. Meituan deployed a quantized 4-bit version of the model on a cluster of 194 Huawei Ascend accelerators from the Atlas A2 series during load testing [ (the-decoder.com) ].
Context
LongCat-2.5 builds on its predecessor, LongCat-2.0, by incorporating native image understanding and improved reinforcement learning based on real software tasks and agent failures [ (the-decoder.com), (cryptobriefing.com) ]. The model achieves competitive performance on several technical benchmarks, posting 88.4 points on Terminal Bench v2.1 and 38.5 on the harder v4.0 variant according to The Decoder [ (the-decoder.com) ]. Meituan has made the model available via an API compatible with OpenAI and Anthropic formats [ (the-decoder.com) ].
What we don't know yet
While previous version weights were released, availability details for the newest iteration's open-weight repositories vary across deployment channels, with some reports noting gradual rollouts to platforms like GitHub and Hugging Face alongside current API access [ (note.com), (technode.com) ]. The long-term scalability of deploying trillion-parameter models entirely on non-Nvidia domestic hardware clusters for broader enterprise workloads remains to be fully seen [ (the-decoder.com), (note.com) ].
Sources
- Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU (the-decoder.com)
- Meituan's LongCat-2.5 Hints at a New Main Battlefield in the Chinese AI War (note.com)
- Meituan releases LongCat-2.5, a 1.6 trillion parameter multimodal model built for agentic work (cryptobriefing.com)
- [LongCat-2.5] What makes Meituan's model so impressive? Dissecting the new star specialized in 1.6T MoE and long-duration agents (note.com)
- Meituan launches LongCat-2.5 multimodal AI model for extended coding tasks (technode.com)
- Meituan Unveils New AI Model Trained on Homegrown Chips (caixinglobal.com)


