Meituan's LongCat-2.5 adds native vision to a 1.6T agent model

The headline change is native vision: LongCat-2.0 was released as a language model[1][3]. Rather than bolting a vision encoder onto the language model, LongCat-2.5 converts images into multi-level discrete tokens using residual vector quantization on native-resolution features, so the backbone processes visual and text input in one token format[1].

Because the image tokenizer is fixed, visual token IDs can be precomputed and cached offline. LongCat says that removes the vision encoder from the training graph and makes it easier to balance image and text workloads across a large cluster[1]. The team pitches the result as visual coding, creative design and 3D work, alongside conventional agentic coding[2].

Most of the post is about reinforcement learning for long tasks. LongCat filters GitHub pull requests and linked issues, has agents build runnable environments from them, and validates each task with fail-to-pass and pass-to-pass tests. It also turns real agent failures into new verifiable training tasks[1].

To curb reward hacking, task environments have no network access, git history and cached files are removed, and an LLM judge audits trajectories for exploits. Extra penalties target repetitive reasoning and repeated tool calls[1].

LongCat's evaluation table compares the model against GPT-5.5, GPT-5.6-Sol, Claude Opus 4.8, Claude Opus 5, GLM-5.3, Kimi K3 and DeepSeek-V4.1-Flash[1].

Benchmark	LongCat-2.5	LongCat-2.0	GPT-5.6-Sol	Claude Opus 5
Terminal Bench v2.1	88.4	70.8	88.8	89.1
Terminal Bench v4.0	38.5	3.0	37.3	53.9
BrowseComp	91.9	79.9	90.4	90.8
MCP Atlas	83.1	61.8	81.8	85.8
AutomationBench	51.8	15.8	45.8	50.3

The jump over LongCat-2.0 is large, especially on the harder Terminal Bench v4.0. Against the frontier it is competitive rather than leading: it edges GPT-5.6-Sol on several agent benchmarks but trails Claude Opus 5 on most coding ones[1]. LongCat notes that competitor scores without an official source were self-run and are indicative only[1].

LongCat says it optimised long-context training and inference for domestic AI accelerators with less than 80 GB of memory each[1]. Its 4-bit weight-and-activation inference path reportedly raised prefill throughput by 43.3% and decode throughput by 42.77% over an 8-bit baseline, with accuracy comparable to the unquantized model on the tasks it evaluated[1].

The launch post does not include downloadable weights, and Meituan's Hugging Face organisation still describes LongCat-2.0 as its latest model[3]. AIbase reported that GitHub and Hugging Face weights will follow[4], but LongCat has not given a date.