Meituan drops 1.6T-parameter multimodal model LongCat-2.5

When people think of Meituan, food delivery couriers on electric scooters usually come to mind. But behind the scenes, the Chinese tech giant has just quietly shipped a trillion-parameter multimodal beast.

Meituan dropped LongCat-2.5, a 1.6-trillion-parameter native multimodal model built specifically for long-horizon agent tasks. It runs on a Mixture-of-Experts (MoE) architecture that activates roughly 48 billion parameters per token. The standout detail: the entire training pipeline ran from start to finish on domestic Chinese hardware.

The model is already accessible through its web chat at longcat.ai, and API access is live for developers. Its GitHub and Hugging Face listings are currently tagged as coming soon, hinting at another open-source release down the line.

Visual coding and long-horizon agents

Unlike traditional multimodal models that patch an external vision encoder onto a text core, LongCat-2.5 treats vision inputs as discrete tokens right alongside text. This native approach pays off in visual coding tasks.

In demonstrations, the model took visual designs and wrote fully interactive websites featuring realistic fluid simulations and physics effects. It can also generate 3D web displays without breaking a sweat.

Where LongCat-2.5 really flexes is in long-horizon software engineering. Meituan pushed sequence lengths up to 1 million tokens during mid-training to handle complex codebases. During reinforcement learning, researchers actually pulled the network plug, wiped git histories, and penalized repetitive tool calls to stop the model from taking shortcuts.

On internal benchmark trackers, Meituan stacks LongCat-2.5 directly against Claude-Opus-4.8 and Claude-Opus-5. The results land comfortably between the two, even outperforming Claude-Opus-5 on several agent-centric evals.

Engineering around hardware constraints

Training a 1.6T MoE model is already brutal. Pulling it off on domestic AI accelerators with under 80GB of VRAM per card requires some serious systems gymnastics.

Meituan restructured its parallelism strategies inside compute super-nodes, ensuring that the bulk of communication stayed on high-bandwidth local fabrics. By precomputing image tokens offline, they stripped the vision encoder out of the live training compute graph to free up precious memory. Technical blog numbers show single-card throughput on long-sequence SFT jumped 86% over standard DDP baselines.

On the serving side, the team implemented a W4A4 quantization setup, pushing both weights and activations down to 4 bits. Compared to standard W8A8 baselines, inference throughput climbed by roughly 40%, while time-to-first-token dropped by 16%.

This fits neatly into CEO Wang Xing's directive from Meituan's recent earnings call: the company will not become a generic "token factory." Instead, LongCat-2.5 powers Miaoshou, Meituan's merchant workspace. In one business analysis case, the model flagged an e-commerce setup where payment GMV grew 23.2% while channel profits dropped 1.7%, diagnosing clearance bottlenecks and recommending specific inventory shifts for major retail chains.

Why it matters

For developers, LongCat-2.5 is living proof that you can train frontier-grade coding agents without relying on Western accelerator silicon. The engineering workarounds Meituan shared around super-node parallelism and W4A4 inference are immediately useful playbooks for anyone building under strict hardware constraints.

If you build software agents or generate web UI from visual mockups, testing the API is definitely worth an afternoon. It excels at multi-step tasks where typical models run out of context or hallucinate halfway through.

Strategically, Meituan's approach looks remarkably sharp. Rather than entering a generic chatbot race, it tuned this model strictly for software automation and real-world merchant operations. Tying AI performance directly to retail unit economics—diagnosing margin drops and managing store inventory—gives Meituan a clear path to monetize compute where others are still burning cash.