Meituan Launches LongCat-2.5, a 1.6-Trillion-Parameter Multimodal Model for Coding Agents
Meituan has released LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts model (with roughly 48 billion active parameters) designed for coding agents and image-based programming, trained entirely on Chinese chips. It is available via an API at $0.30 per million input tokens and $1.20 per million output tokens (promotional pricing), with no open-weights release announced so far.
Chinese delivery giant Meituan has launched LongCat-2.5, its new flagship large language model. It is a mixture-of-experts model featuring 1.6 trillion parameters, of which around 48 billion are active for any given token. It is engineered for coding agents that autonomously handle extended tasks, as well as for image-based programming. Meituan emphasizes that it was trained entirely on Chinese chips. The model is available on longcat.ai and via API, at a promotional rate of $0.30 per million input tokens and $1.20 per million output tokens. No open-weights release has been announced at this time.
What Changed Since the Preview
LongCat-2.5 is the final production version of LongCat-2.5-Preview, which went live on September 25. The engineering team’s blog post is dated November 1, and the API changelog is dated November 2. The model builds on the foundation of LongCat-2.0, released in June, focusing on two main areas of development.
The first is native vision. Images are converted into discrete tokens, much like words, and processed by the same underlying model as text without a separate vision encoder. Meituan leverages this for “visual programming”: reconstructing web pages from design mockups, generating interactive 3D pages, or building data visualizations.
The second is long-horizon agentic execution. The team trained the model using reinforcement learning on tasks derived from GitHub pull requests and validated by test suites. They also report turning real-world agent failures into new training exercises. To prevent cheating, network access is cut off during these tasks, and Git commit histories are stripped out. The changelog notes out-of-the-box compatibility with Claude Code, Codex, and Hermes.
Benchmark Scores Rivaling Claude Opus 5, According to Meituan
The blog post includes a benchmark table covering roughly thirty evaluations. According to these figures, LongCat-2.5 scores 88.4 on Terminal Bench v2.1 compared to 89.1 for Claude Opus 5, and 91.9 on BrowseComp compared to 90.8. It lags significantly behind on Terminal Bench v4.0, posting 38.5 versus 53.9. Meituan’s press release summarizes its performance as landing between Claude Opus 4.8 and Claude Opus 5, edging out Opus 5 on a subset of the tests. On several coding benchmarks, DeepSeek-V4.1-Flash still holds the lead, posting 90.6 on Terminal Bench v2.1 and 74.2 on DeepSWE, compared to 64.6 for LongCat-2.5.
The Domestic Chip Factor
Meituan places heavy emphasis on the hardware side of the release. Its chips have less than 80 GB of memory each, which required rearchitecting its distributed computing setup. According to the blog, per-chip training throughput improved by 86% on long-context tasks. For inference serving, expert weights and activations are quantized to 4-bit precision. Published stress tests indicate this provides roughly a 43% throughput increase over the 8-bit version.
Our take
LongCat-2.5 confirms that major Chinese tech firms outside traditional AI research labs—in this case, an on-demand food delivery giant—are capable of training top-tier agent models without relying on American silicon. That remains the most significant takeaway, regardless of the benchmark scores. Those numbers should also be taken with a grain of salt: they are published by Meituan, some competitor baselines were measured internally, and the comparison stops at Claude Opus 5 and GPT-5.6 Sol, even though newer models are already in production. The introductory pricing is aggressive, but explicitly framed as temporary. For developers, the lack of open weights is the main drawback; Meituan open-sourced several previous LongCat models, but without access to the weights here, running the model locally or independently verifying its claims remains impossible.
Sources
- LongCat (Meituan) — Introducing LongCat-2.5 Primary source
- LongCat API Docs — Changelog (Release 2026-11-02) Primary source
- LongCat API Docs — LongCat-2.5 Pricing Primary source
- Meituan — 美团发布LongCat-2.5大模型 (Press release via The Economic Observer) Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
