LongCat 2.5 released

LongCat 2.5 is an open-weight model from Meituan built for coding agents and other long, multi-step tasks, and Meituan's own benchmark table puts it far ahead of its predecessor and in the same range as the frontier scores it cites, ahead on some tests and behind on others. The weights on the Hugging Face model card are released under the MIT license, which allows commercial use and modification. The model reads images as well as text: the card says images are converted into discrete tokens that share a representation with text, so one model handles both.

LongCat 2.5 is a mixture-of-experts (MoE) model. Instead of one block of parameters that handles every input, each layer holds many smaller sub-networks, called experts, and a router sends each token (a word or piece of a word) to a few of them. The card gives 1.6 trillion parameters in total, with about 48 billion active for each token, so the computing work per token is far smaller than the total size suggests, but all 1.6 trillion parameters still have to be held in memory. This repository stores them in BF16, a 16-bit format, and Hugging Face's file listing puts the weight files at about 3.6 terabytes, far more than a single accelerator chip can hold. The only deployment route the card documents is for NPU platforms, NPU being a general term for chips built for AI workloads. Meituan also published a 4-bit version for cheaper serving, and the model can be called without self-hosting through Vercel's AI Gateway.

In the card's table, LongCat 2.5 scores 88.4 on Terminal Bench v2.1, a test of completing tasks in a command-line environment, up from 70.8 for LongCat-2.0, the same score Nous Research cited when it offered the earlier model. On that test the table lists GPT-5.6-Sol, Claude-Opus-5 and DeepSeek-V4.1-Flash between 88.8 and 90.6, with GLM-5.3 and kimi-k3 just below LongCat 2.5 at 88.2 and 88.3. Across the ten coding-agent rows, LongCat 2.5 is below the Claude-Opus-5 figure on eight. Meituan credits the model's abilities to long-sequence training, large-scale reinforcement learning on real engineering tasks and reproducible agent failures, training guided by several specialized teacher models, and infrastructure work. The card does not show how much each of these contributed to the gains over LongCat-2.0.

The comparisons need care. The card marks scores cited from public technical reports, blog posts or official leaderboards with an asterisk and calls its own in-house measurements "indicative only", and many of its coding tests ran inside the Claude Code agent harness under settings Meituan chose. A row in the table can therefore set a LongCat score measured one way beside a rival's score measured another way. What the release establishes is that a very large, permissively licensed agent model is now downloadable. How close it really comes to the leading closed models awaits independent testing.