Meituan

Native multimodal foundation model (74B MoE, ~3B active) unifying text, vision, and audio understanding and generation via DiNA (Discrete Native Autor…

560B MoE model for formal theorem proving in Lean4. Achieved 97.1% pass rate on MiniF2F-Test.

Introduces a principled framework for Batch Size Scheduling (BSS) based on functional scaling laws. The paper uncovers the "fast catch-up" effect, sho…

Smaller, cost-efficient LongCat variant based on the "Scaling Embeddings Outperforms Scaling Experts" research. Designed for high-throughput productio…

Updated 560B MoE agentic reasoning model with DORA RL framework and Heavy Thinking mode. Achieves 97.1% on theorem proving benchmarks. A distinct rele…

"Evolving Computer Use Agent" — framework for training agents to navigate software and web interfaces.

Introduces LoZA (LongCat ZigZag Attention), a sparse attention scheme for efficient long-context scaling in LongCat models.

6B parameter bilingual (Chinese-English) image generation model outperforming larger open-source models. Includes editing and a 10x speedup distilled…

Native omni-modal model supporting streaming audio-visual interaction. 560B MoE (27B active), 128K context, millisecond-level end-to-end latency, 8+ m…

13.6B parameter foundational video generation model unifying text-to-video, image-to-video, and video-continuation tasks. Uses Diffusion Transformer w…

Audio tokenizer and detokenizer designed for speech LLMs. Underpins the LongCat-Flash-Omni model's audio capabilities.

Proposes RITE, a reinforcement learning framework for training LLM agents that interleave reasoning with tool use across diverse domains. Enforces con…

Dataset for evaluating agentic web search and synthesis.

Introduces CoTP, a framework that selects high-value Chain-of-Thought (CoT) data by abstracting atomic reasoning patterns. Using only 10B tokens of Co…

Reasoning variant of LongCat-Flash trained via large-scale RL. Rivals proprietary systems in complex logic and STEM benchmarks.

560B parameter MoE model activating ~27B per token. Meituan's foundational LLM with PID-controller-based dynamic expert allocation and "Zero-computati…

Benchmarks for evaluating visual and logical reasoning.

Studies how LLMs can use self-generated tests to debug their own code, comparing two paradigms: post-execution self-debugging (generate code, run test…

An agent synthesis framework that enables models to learn from observations within trajectories to autonomously correct errors. Accepted to ICLR 2025,…

A transparent, step-by-step guide to estimating the constant terms in scaling law formulas. By training small models (1M-60M parameters), the authors…

"Language Enhanced Reasoning Detection Assistant" combining MLLMs with object detection for reasoning-based detection.

Benchmark for reasoning-based object detection.

High-performance single-stage object detector dedicated to industrial applications. One of Meituan's most famous open-source contributions.

AI navigation and safety stack powering Meituan's fleet of unmanned delivery drones and L4-level autonomous ground robots. Completed hundreds of thous…