Meituan, the Chinese delivery company, has released LongCat-2.5, which its launch blog calls “a native multimodal model for long-horizon agentic tasks”. It has 1.6 trillion total parameters with about 48 billion active per token, and builds on LongCat-2.0.
What is new
- Images as tokens. LongCat-2.5 turns visual input into multi-level discrete tokens, using residual vector quantization on native-resolution features, so the same backbone handles images and text. Meituan pitches this for visual coding, design and 3D tasks.
- Long agent runs. Meituan says it trained on more long-horizon trajectories with large-scale reinforcement learning, and turns real agent failures into verifiable training tasks.
- Home-grown chips. Meituan says it trained long-context stages and runs 4-bit weight and activation (W4A4) inference on “domestic AI accelerators” with under 80 GB per device. It does not name the vendor. In its own stress tests, W4A4 raised prefill throughput by 43.3% and decode throughput by 42.77% over an 8-bit baseline.
The numbers, with caveats
All scores are Meituan’s. Rival scores marked with an asterisk in its table come from public reports; the rest were run by Meituan and are, in its words, “indicative only”. A few starred comparisons:
| Benchmark | LongCat-2.5 | Claude Opus 5 | Other |
|---|---|---|---|
| Terminal Bench v4.0 | 38.5 | 53.9 | GPT-5.6-Sol 37.3 |
| DeepSWE v1.1 | 64.6 | 68.8 | DeepSeek-V4.1-Flash 74.2 |
| BrowseComp | 91.9 | 90.8 | GPT-5.6-Sol 90.4 |
So it leads on some agent and browsing tests and trails Opus 5 and DeepSeek’s Flash on several coding ones. Meituan ran most coding tests in the Claude Code harness.
Price and availability
LongCat-2.5 is on Meituan’s API pay-as-you-go plan at a “limited-time” discounted price of $0.30 per million uncached input tokens, $0.006 cached and $1.20 output. The page says prices are subject to change.
There are no weights. As of the evening of 2 November, Meituan’s Hugging Face page had no LongCat-2.5 repository, and the blog does not promise one.
Why this matters
If you route agent work by price, LongCat-2.5 is worth a test on your own tasks while the promo lasts, but plan for the price to change. Unlike DeepSeek, which posted V4.1-Flash weights at launch, or Moonshot, which promised Kimi K3.1 Pro weights by a date, Meituan offers API access only, so you cannot self-host it today. Check your own data rules before sending code or images to a new provider.
The Campfire
CommentsThanks. Your comment is waiting for the desk. Once it is approved it appears here within about a minute.
Add a comment