Meituan's LongCat-2.5 Takes On Claude Opus 5 With 1.6 Trillion Parameters

Meituan has released LongCat-2.5, a native multimodal AI model with 1.6 trillion total parameters that the Chinese company says scores between Anthropic's Claude Opus 4.8 and Claude Opus 5 on most mainstream benchmarks. The model, announced on November 2, activates about 48 billion parameters per token and is built for long-running agent tasks such as working through large codebases and building complete software projects.

It matters for developers because LongCat-2.5 is pitched directly at the coding-agent workflows many teams already use. Meituan says it works with harnesses including Claude Code, Codex and Hermes, and it is available now through the company's chat service and API. It is also a test case for whether a frontier-scale model trained entirely on Chinese-made accelerators can compete with the leading US systems.

Key facts at a glance

• Size: 1.6 trillion total parameters in a mixture-of-experts (MoE) design, with roughly 48 billion active for each token.
• Focus: Long-horizon agentic coding plus native visual understanding, including visual coding, design and 3D tasks.
• Hardware: Meituan says training was done entirely on domestic Chinese compute, using accelerators with less than 80 GB of memory each.
• Access: Available via the LongCat chat site and API from launch; GitHub and Hugging Face releases are planned to follow.
• Benchmarks: 88.4 on Terminal Bench v2.1, against 89.1 for Claude Opus 5 and 88.8 for GPT-5.6-Sol in Meituan's comparison table.

How LongCat got here

LongCat is Meituan's in-house model family. The delivery and local services giant released LongCat-2.0 in June, which it described as the first time it had completed full training and inference of a trillion-parameter model on domestic hardware. A LongCat-2.5-Preview followed on September 25, adding image understanding and tighter integration with coding tools.

The company is clear about why it builds models. On its second-quarter earnings call on August 28, chief executive Wang Xing said Meituan does not intend to become a "token factory" and that its models and AI products are meant to support its core business. In line with that, Meituan says LongCat-2.5 was strengthened for business analysis and software development, and it already powers an AI operations workbench called Miaoshou that serves restaurant and drink chains in China.

For developers outside China, the more relevant point is that LongCat is increasingly positioned as a general-purpose coding agent model that slots into Western tooling rather than a domestic-only product.

What's new in LongCat-2.5

Native multimodality. Rather than bolting a separate vision encoder onto a language model, LongCat-2.5 turns images into multi-level discrete tokens using residual vector quantisation, so the backbone handles pictures and text in one token format. Meituan says this lets it precompute and cache image tokens offline, which simplified large-scale training, and supports tasks such as turning a design into an interactive web page or 3D exhibit.

Agent training at scale. The team says it added more long-horizon task trajectories, used large-scale reinforcement learning for end-to-end engineering work, and built a pipeline that mines GitHub pull requests and linked issues to create training tasks with automatic tests. It also turns real failed agent runs into reproducible training tasks to patch capability gaps.

Guarding against reward hacking. Meituan says it disabled network access in training environments, removed git history and cached files, and used another model as a judge to audit trajectories for shortcuts. Extra penalties were added for repetitive reasoning and repeated tool calls, a common failure in long agent sessions.

Efficiency on local chips. In mid-training the model used sequences of up to 1 million tokens. Meituan reports its sharded training setup delivered 86% higher per-device throughput than a standard data-parallel baseline on long-sequence fine-tuning. For serving, it quantises expert weights and activations to 4 bits (W4A4). Against an 8-bit baseline, the company reports 43.3% higher prefill throughput, 42.77% higher decode throughput, a 15.91% cut in time to first token and a 19.72% cut in time per output token.

How it compares on benchmarks

Meituan published a broad table comparing LongCat-2.5 with OpenAI, Anthropic and Chinese rivals. A selection of agent and coding results:

Figures as published by Meituan in its LongCat-2.5 technical post.

The picture is mixed rather than dominant. LongCat-2.5 edges ahead on web browsing research (BrowseComp) and roughly matches the leaders on the older Terminal Bench, but trails Claude Opus 5 clearly on the harder Terminal Bench v4.0 and on Program Bench, where it scores 25.5 against 37.0. The jump from LongCat-2.0 is large: on Terminal Bench v4.0 the older model scored 3.0.

The counterpoint: read vendor benchmarks carefully

These are company-reported numbers. Meituan says scores for other models marked as taken from public reports or leaderboards are reused, while the rest were run by its own team and are "indicative only." Many results also depend on the harness and settings used; Meituan notes, for example, that several coding tests ran in Claude Code with a 512K token sequence limit.

There is also a gap between a model being available and being practical. At launch, every LongCat-2.5 score in the table came from Meituan itself, and the release on GitHub and Hugging Face has been announced but not yet delivered. Until it is, teams can only use LongCat-2.5 through Meituan's hosted service.

What it means for you

For developers in the US, UK, Canada and Australia, LongCat-2.5 is another option for agentic coding that plugs into tools such as Claude Code and Codex, which can make side-by-side testing relatively cheap. Its strengths, according to Meituan's own numbers, are browsing-heavy research tasks and visual coding from designs or screenshots.

Before routing work to it, check where your prompts and code will be processed and under what terms, as you would for any hosted model. Teams handling client code or regulated data should review the provider's data handling and retention terms and their own contractual obligations first.

For investors and industry watchers, the bigger signal is hardware. If Meituan's efficiency claims for training and 4-bit inference on Chinese accelerators hold up, they would add to evidence that limits on access to advanced US chips have not stopped frontier-scale model development in China.

What happens next

The next milestone is the promised release on GitHub and Hugging Face, which will show whether weights are open and under what licence. Independent benchmark runs on public leaderboards will follow from that. Developers should also watch API pricing and rate limits once launch promotions end.

Takeaways

• Test it on your own tasks: run LongCat-2.5 through the same harness and repositories you use for Claude or GPT models before drawing conclusions.
• Weigh the hard benchmarks most: its gap on Terminal Bench v4.0 and Program Bench matters more for complex projects than near-ties on older tests.
• Check data terms first before sending proprietary code to any hosted model.
• Watch for the open release, which would make local or self-hosted evaluation possible.
• Track the hardware story: 4-bit inference and training on chips under 80 GB are claims worth verifying as independent results appear.