LongCat-2.5
modelYour notes
The successor to LongCat-2.0, released as an API model in two steps. Meituan's LongCat API Platform change log lists LongCat-2.5-Preview on 2026-09-25 ("New image understanding capability, able to parse image content, supporting cross-modal Q&A, content summarization, and complex visual reasoning") and the general release, LongCat-2.5, on 2026-11-02. The GA entry names three features: "Native Multimodality, Beyond Text" (visual understanding combined with agentic coding for visual coding, creative design and 3D tasks), "Long-Horizon Agents for Complex Engineering" (complex codebases and full-project builds), and compatibility across agent harnesses (Claude Code, Codex, Hermes and others). The main change from 2.0, which was text-only, is native image input.
The platform documents a 1M-token context window with a 256K-token maximum output, and LongCat-2.5 is now the only model on the platform. Limited-time pricing is $0.30 per 1M uncached input tokens, $0.006 cached and $1.20 output (¥2 / ¥0.04 / ¥8).
Open weights followed on 2026-11-04, two days after the API release, under the MIT license on HuggingFace, with a separate W4A4 checkpoint. The model card confirms the shape the Chinese press reported for the Preview: a native multimodal MoE with 1.6T total and about 48B active parameters, "building on LongCat-2.0". Its config.json gives 38 layers at hidden size 8192, MLA attention with a sparse top-2048 indexer, 768 routed experts at top-12 plus 128 identity "zero" experts, and a 160K vocabulary. The card does not say whether 2.5 continues 2.0's pretraining or how many tokens were added. It describes the training at a high level:
- Vision: native-resolution visual features are discretized into tokens and aligned with text in a shared space, following LongCat-Next.
- Post-training: long-horizon agentic RL on real engineering tasks and reproduced agent failures, plus multi-teacher full-vocabulary KL distillation from agent, reasoning and interaction teachers.
- Infrastructure: trained and served on "domestic AI accelerators". Supernode-aware parallelism gave 86% higher per-device throughput than a DDP baseline in long-sequence SFT, and training and inference produce bitwise-identical token log-probabilities.
- Quantization: W4A4 on expert weights and activations with FlatQuant-style learned transforms. Against W8A8 it reports 43.3% higher prefill and 42.77% higher decode throughput, and W4A4KV8 gives 92% higher decode throughput than W8A8KV16. Deployment instructions target NPUs through Meituan's SGLang-FluentLLM fork.
Self-reported results, with thinking on: Terminal Bench v2.1 88.4, Terminal Bench v4.0 38.5 (2.0: 3.0), DeepSWE v1.1 64.6, BrowseComp 91.9, HLE with tools 63.1, MCP Atlas 83.1, PaperBench 82.1, GDPval-AA rubrics 84.8 and Vision2Web 78.6. In Meituan's table it trails Claude Opus 5 on most coding rows (Terminal Bench v4.0 53.9, DeepSWE 68.8) and leads GLM-5.3 on several agent rows. Unstarred scores are in-house runs that Meituan calls "indicative only". There is no technical report yet.
Model Details
Variants
| Name | Parameters | Notes |
|---|---|---|
| LongCat-2.5 | 1.6T | Full-precision weights |
| LongCat-2.5-W4A4 | 1.6T | W4A4 expert quantization for INT4 on NPUs |