Meituan launches LongCat-2.5, a 1.6-trillion-parameter model for coding agents

Meituan, the Chinese giant best known for delivery and local services, launched LongCat-2.5 on November 2, a new multimodal AI model from its LongCat lineup. According to TechNode, which cites an original report from the Chinese site IT Home, the model has 1.6 trillion total parameters but activates around 48 billion per processed token. The company's stated goal is clear: to sustain AI agents in coding tasks that require prolonged execution, not just one-off responses.

A giant that only wakes up a fraction of itself at a time

The design with 1.6 trillion parameters and only 48 billion activated per token is a Mixture of Experts (MoE) architecture: the model holds a large amount of accumulated knowledge, but only switches on a small slice of it for each inference. This is what makes it feasible to run a model this size without an inference cost proportional to the total parameter count. It's the same logic behind other recent Chinese MoE models, such as DeepSeek V3 (671 billion total parameters, ~37 billion activated), which helped popularize this architecture outside Western labs.

LongCat-2.5's activated share is about 3% of the total, close to the range used by other leading MoE models. In practice, for those consuming it via API, what matters is the cost per processed token, not the total model size stored on Meituan's servers. The source doesn't provide per-token pricing or performance benchmarks, so any real cost comparison still depends on independent testing.

Agents that don't break mid-task

The central angle of Meituan's announcement isn't just size: it's durability in long tasks. The company says it expanded training on extended task sequences and used large-scale reinforcement learning to improve performance on complex codebases and complete software projects, not just isolated snippets of code.

One practical detail of the process stands out: Meituan says it converts real agent failures into reproducible training tasks. In other words, when an agent makes a mistake in production (or in testing), that error becomes training material to close the gap that caused the failure. It's a feedback loop similar to what MLops teams do manually with error logs, except it's built into the model's own training pipeline.

For those building with autonomous coding agents, this is directly relevant: the most common pain point today isn't generating an isolated function, it's maintaining context and coherence across a task that spans dozens of steps, opens multiple files, and has to handle errors along the way without restarting from scratch.

Multimodality without multiple separate models

LongCat-2.5 processes visual and text inputs within a single token framework, according to Meituan. This means image and text are handled by the same internal mechanism, rather than going through separate modules stitched together afterward. The company cites three concrete uses for this capability:

• Visual coding, useful for generating or adjusting interfaces from screens or wireframes;
• Creative design, aimed at tasks that mix visual content generation and text;
• 3D tasks, a category the source mentions without detailing examples.

None of these three uses comes with a public demo, video, or benchmark in the available material. These are capabilities announced by the company, not independently verified so far.

Where it can already be used, and what's still missing

LongCat-2.5 has been available since launch on Meituan's online chat service and on the company's API platform. Anyone looking to integrate the model into a product or agent workflow can already do so via API, along the same lines as what other Chinese companies (DeepSeek, Alibaba with Qwen, Moonshot with Kimi) have been offering.

What doesn't exist yet is an open release: Meituan says GitHub and Hugging Face releases are planned to follow the API launch, but they hadn't been published as of the report's date. This is an important difference from models like DeepSeek V3 and Qwen, which made their weights open from the start or shortly after the announcement. Until that release happens, running LongCat-2.5 locally or auditing its weights isn't an option.

What changes for developers in Brazil

LongCat-2.5 enters a race that was already crowded: Chinese open-weight (or soon-to-be-open) models specifically tuned for agentic coding, competing on inference cost with closed Western models. For Brazilian teams evaluating which model to use behind a coding agent, this means one more option to test once the weights are released, but no concrete performance numbers to compare today.

Two practical points are worth noting before adopting it: first, the API runs on Meituan's infrastructure, which raises the same data residency and privacy questions already discussed with other Chinese AI providers; second, the company didn't publish any benchmark (SWE-bench, HumanEval, or equivalent) alongside the announcement, so any claim of gains in long coding tasks still depends on independent evaluation once the model (or its weights) becomes accessible for testing.

The real validation of LongCat-2.5 for developers will only come when it's tested outside Meituan's own controlled environment, whether via API in real projects, or with open weights on GitHub and Hugging Face once that release happens.