Meituan launches LongCat-2.5 to tackle long-running software projects
The Chinese tech giant just dropped a multimodal AI model trained specifically to execute extended coding tasks without losing the plot.
Models · Source: TechNode
What happened
Meituan released LongCat-2.5 on November 2. It is a native multimodal AI model built for extended coding tasks. Instead of just spitting out code snippets, it is designed to execute long sequences of actions across complex codebases. The model is available now via the LongCat API and chat interface. GitHub and Hugging Face releases are planned to follow.
The training process focused heavily on long-horizon reinforcement learning. Meituan fed the model extended sequences of agent actions. They took failed runs from real AI agents and turned them into reproducible training tasks. The system receives feedback on the outcome of a full sequence rather than treating each response as a separate exercise. This teaches the model how to recover when a software project goes off the rails.
The model processes text and visual inputs in a unified token framework. It converts images into tokens at multiple detail levels. Meituan claims this supports visual coding, creative design, and 3D tasks. The company also optimized the model using 4-bit weights and activations. This W4A4 inference optimization is specifically designed to run these long tasks efficiently on domestic Chinese chips.
Key facts
- 88.4 — Reported score on Terminal Bench v2.1
- 1.6 trillion — Reported total parameters, though this figure is disputed
- W4A4 — Inference optimization used for domestic chips
Why it matters
We are moving past the era of AI as a simple autocomplete tool. Builders need agents that can navigate a messy codebase and see a feature through to completion. By training on long sequences and failed attempts, Meituan is trying to solve a massive bottleneck. They are targeting the exact problem founders face when AI agents get confused halfway through a complex engineering task.
The hardware optimization is a massive signal for the global AI race. Meituan optimized W4A4 inference specifically for domestic chips. This means Chinese tech giants are actively building workarounds for hardware export constraints. They are ensuring their long-running agents can operate at scale without relying entirely on the latest Western silicon. This forces everyone else to rethink how efficiently agentic workflows can run.
For builders
Plug into existing agent harnesses
LongCat-2.5 already works with Claude Code, Codex, and Hermes. Meituan also provides setup guides for Cline and WorkBuddy. Engineers can swap this model into their current agent workflows today to test its endurance.
Leverage visual coding capabilities
The model handles images and text in a unified token framework. Founders building UI generation or design-to-code tools can use this to process visual inputs directly. It opens up new ways to automate frontend development.
Watch out for disputed parameter counts
Reports claim the model has 1.6 trillion total parameters with 48 billion active per token. However, earlier reports dispute this, suggesting these numbers were just copied from the older LongCat-2.0 model card. Do not base your infrastructure costs on these unverified specs.
My take
A food delivery company is shipping AI models that try to build entire software projects. That is the reality of the market right now. If you are still building wrappers for single-shot prompts, you are going to get crushed by agents that actually have the patience to finish the job.
Original reporting: TechNode. This is my rewrite and opinion.