SSyncDev

LongCat-2.5

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

LongCat-2.5 is Meituan's open-weight MIT-licensed multimodal MoE for long-horizon agents: screenshots, charts, video and multi-step coding.

Key specifications

Specification score
30.0
License
MIT

The specification score summarizes recorded features and limits. It is not a benchmark result.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input
  • Image input
  • Video input

Produces

  • Text output

Capabilities

ReasoningVisionOpen weights

What is LongCat-2.5?

Can an open model look at a broken web page, find the bug in the repository and patch it? That is the kind of loop Meituan built this release for. It is a native multimodal mixture-of-experts aimed at long-horizon agentic work, reading screenshots, charts, images and video alongside text and then writing or revising code over many steps.

The final weights arrived on 9 November 2026 under the MIT licence, following LongCat-2.5-Preview in September.

Specifications

Item Detail
Hugging Face repo meituan-longcat/LongCat-2.5
Parameters 1.6T total, about 48B active (file metadata reads 1.8T)
Input Text, image, video
Output Text
Modes Thinking and non-thinking, via enable_thinking
Quantisation W4A4 for expert weights and activations
Licence MIT (no trademark or patent rights)
Chat longcat.ai

Context length is not on the card

The model card does not state a native context window. Evaluation footnotes use maximum sequence lengths of 256K, 512K and 1M for different benchmarks. Until Meituan publishes a figure, test the length you need instead of assuming 1M.

Deployment is NPU-first

The published path runs on domestic AI accelerators through the NPU branch of SGLang-FluentLLM. No GPU or vLLM recipe is given on the card, which matters if your cluster is NVIDIA-only. W4A4 quantisation is the efficiency lever: against a W8A8 baseline Meituan reports about 43% higher prefill throughput and 43% higher decode throughput.

With thinking on, Meituan recommends keeping prior reasoning in history (save_history_reasoning_content=True). Thinking off trades some quality for token efficiency.

How it moved on from 2.0

The release builds on LongCat-2.0 with native visual modelling, long-horizon agentic reinforcement learning and multi-teacher learning. Publisher-reported scores show the jump: Terminal Bench v2.1 rises from 70.8 to 88.4, BrowseComp from 79.9 to 91.9, and DeepSWE v1.1 from 9.7 to 64.6. Meituan does not lead every comparison it publishes.

A test that matches the pitch

Give it a failing UI screenshot, a test repository and a way to render the patched page. Then check whether it finds the cause and makes a targeted change rather than a sprawling rewrite. Run the same task with thinking on and off, and note tokens used. If you only need a lighter model, the Meituan provider page lists the rest of the family.

How to evaluate LongCat-2.5

Compare the model on the work you need to complete.

Tasks and inputs

LongCat-2.5 is listed for Image-text-to-text, Multimodal, Video-text-to-text, Coding & Code Generation. It accepts text, image, video and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under MIT. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

The weights are released under the MIT licence, which grants no rights to Meituan trademarks or patents.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

sensetime

SenseNova-Loop-9B

SenseNova-Loop-9B is SenseTime's 9B looped multimodal model: 32 layers run as 56, 512K context, ~15 GiB at 256K in 4-bit, open weights under Apache 2.0.

524.29K context

Tencent

Hy4 Flash

Tencent Hy4 Flash is the low-cost Hy4 tier with open weights: a 220B MoE (~11.3B active) for agents and documents, with 1M context and image input.

1.05M context

shanghai-ai-lab

Sage S2.1 Flash

Sage S2.1 Flash is Shanghai AI Lab's open scientific multimodal agent model, reading molecules, crystals and proteins alongside text and images.

— context

sensetime

SenseNova-Loop-9B-Preview

SenseNova-Loop-9B-Preview is SenseTime's looped 9B text-and-image model for long-context agentic work, running shared layers twice for extra depth.

262.14K context

Liquid AI

LFM3-VL-2B

LFM3-VL-2B is Liquid AI's open 2B vision-language model for images and video on-device, strong on documents, charts and GUI screenshots.

— context

NVIDIA

Nemotron 3.5 Super VL 120B A12B

Nemotron 3.5 Super VL 120B A12B is NVIDIA's open multimodal reasoning MoE: 12B active, 1M context, image and video input, switchable thinking.

1.05M context