AIsaAIsaStudio 中文 Open AIsa Studio

Meituan LongCat releases LongCat-2.5, a native multimodal MoE model with 1.6T total parameters

Original title: meituan-longcat/LongCat-2.5 (image-text-to-text)

Meituan's LongCat team released LongCat-2.5 on Hugging Face, a native multimodal MoE model for long-horizon agentic tasks with 1.6 trillion total parameters and about 48 billion activated per token, under the MIT license. Building on LongCat-2.0, it combines visual understanding with agentic coding; training uses long-sequence SFT, large-scale RL, and multi-teacher KL distillation, with reported per-device throughput 86% higher than the DDP baseline in long-sequence SFT tasks. Benchmarks show gains over LongCat-2.0, e.g. Terminal Bench v2.1 at 88.4 vs 70.8.

Why it matters

The README details parameter scale, training methods, and benchmark comparisons against several frontier models, letting readers gauge where it stands on agentic coding tasks.

Read the original ↗

The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.

Original article

Hugging Face 热门模型第 29 名:meituan-longcat/LongCat-2.5,241 赞,近期下载 971 次。 标签:safetensors, pro_v, multimodal, image-text-to-text, text-generation, mixture-of-experts, agentic, conversational, custom_code, arxiv:2603.27538, arxiv:1909.07606, license:mit

README:

LongCat-2.5

Tech Blog 📄

Model Introduction

LongCat-2.5 is a native multimodal MoE model for long-horizon agentic tasks, with 1.6 trillion total parameters and approximately 48 billion activated per token. Building on LongCat-2.0, it combines visual understanding with agentic coding for complex codebase development, full-project builds, creative design, and 3D tasks. Long-sequence training, large-scale RL, multi-teacher learning, and infrastructure optimizations on domestic AI accelerators underpin these capabilities.

LongCat-2.5 Highlights

  • Native Visual Modeling: Building on LongCat-Next, native-resolution visual features are converted into discrete tokens and jointly aligned with text in a shared semantic space.

  • Long-Horizon Agentic RL: Training combines real-world engineering tasks and reproducible agent failures with efficient rollout strategies and safeguards against reward hacking, improving reliable end-to-end task execution.

  • Multi-Teacher Learning: Specialized agent, reasoning, and interaction teachers guide the model through full-vocabulary KL distillation and Tree Attention, strengthening multi-turn execution, reasoning, instruction following, and safety.

  • Efficient Long-Context Training: Supernode-aware parallelism, efficient sharding, and communication–computation overlap improve training efficiency, achieving 86% higher per-device throughput than the DDP baseline in reported long-sequence SFT tasks.

  • RL Infrastructure: Drawing on community practices, aligned training and inference logic maintains bitwise-identical token log probabilities across training, prefill, and decoding. KV cache optimization and LSA improve serving efficiency, nearly doubling decode KV capacity without increasing TPOT on the tested workload.

For full technical details, see our blog.

Evaluation Results

We evaluate LongCat-2.5 across coding, general agent, and visual tasks. Scores marked with * are cited from public technical reports, blog posts, or official leaderboards; unmarked scores are measured in-house and are indicative only. - indicates an unreported score. See the expandable footnotes below for evaluation settings.

BenchmarkLongCat-2.5LongCat-2.0GPT-5.5GPT-5.6-SolClaude-Opus-4.8Claude-Opus-5GLM-5.3kimi-k3DeepSeek-V4.1-Flash

Code Agent Terminal Bench v2.188.470.883.288.874.689.188.288.390.6* Terminal Bench v4.038.53.0-37.323.653.941.812.631.2 Program Bench (Almost Resolved)25.52.513.523.0<t

The original belongs to Hugging Face 热门模型 and its authors and is reproduced here for reading. To have it removed, write to support@aisa.one.