Meituan (LongCat)

{"content":"## Thesis\n\nMeituan's LongCat lab is a full-stack frontier AI organization: it ships frontier-scale MoE language models (LongCat-2.0 at 1.6T total / ~48B activated parameters) while simultaneously building the inference systems, sparse-attention kernels, and evaluation suites those models require P5P6P24. Its most distinctive strategic bet is compute sovereignty — LongCat-2.0's training and deployment run entirely on AI ASIC / domestic "国产" compute clusters, claimed as the industry's first trillion-parameter model trained and served end-to-end on ~50k domestic accelerator cards P5W1W6. Around that core, the lab spans multimodal generation (video, image, audio, omni), reasoning/formal-math models, agentic deep-research systems, and an unusually broad public benchmark portfolio (VitaBench, AMO-Bench, Meeseeks, R-HORIZON, UNO-Bench, MineExplorer, SOP-Maze, PIE_bench, WBench, General365, LARYBench) P13P17P8P14P15P7P12P21E32E38E34. The pattern is "model–system co-design": LongCat releases models alongside the kernels and engines needed to serve them efficiently P24P3P4.

Signal desks\n\nHiring\n- org=meituan-longcat · team=LongCat large-model team · location=global (full-time master's/PhD students worldwide) · role_theme=AGI talent pipeline via a 2026 large-model internship program; focus areas are base-model algorithms, infrastructure architecture, agent applications, and visual generation · evidence_url=https://www.xix.ai/es/ainews/meituan-unveils-longcat-ai-model-internship-program-to-cultivate-agi-talent.html W5. Note: this is the only hiring signal in the pack — no full-time role listings or org-chart data are cited.\n\nForks\n- org=meituan-longcat · upstream=deepseek-ai/FlashMLA · language=— · theme=MLA attention kernel (SwapAB / FP8 KVCache) · evidence_url=https://github.com/meituan-longcat/FlashMLA E52P24\n- org=meituan-longcat · upstream=deepseek-ai/DeepGEMM · theme=GEMM kernel (SwapAB Offset + PDL) · evidence_url=https://github.com/meituan-longcat/DeepGEMM E56P24\n- org=meituan-longcat · upstream=deepseek-ai/DeepEP · theme=expert-parallelism communication · evidence_url=https://github.com/meituan-longcat/DeepEP E57\n- org=meituan-longcat · upstream=flashinfer-ai/flashinfer · theme=attention/inference kernels (communication–computation fusion) · evidence_url=https://github.com/meituan-longcat/flashinfer E58P24\n- org=meituan-longcat · upstream=microsoft/mscclpp · theme=multi-GPU/NPU collective communication · evidence_url=https://github.com/meituan-longcat/mscclpp E49\n- org=meituan-longcat · upstream=Dao-AILab/fast-hadamard-transform · theme=Hadamard/rotation kernel (quantization-adjacent) · evidence_url=https://github.com/meituan-longcat/fast-hadamard-transform E54\n\nThese forks cluster on 2026-02-05/06, immediately around the SGLang-FluentLLM engine launch (2026-02-03), signaling a kernel-level inference push rather than general curiosity E52E56E58E36.\n\nReleases\n- org=meituan-longcat · release_type=model weights (HF) · artifact=LongCat-2.0 (1.6T/48B MoE, text-gen) · traction=1,728 downloads / 273 likes · evidence_url=https://huggingface.co/meituan-longcat/LongCat-2.0 E3P5\n- org=meituan-longcat · release_type=model weights · artifact=LongCat-Flash-Chat (560B) · traction=23,871 downloads / 537 likes · evidence_url=https://huggingface.co/meituan-longcat/LongCat-Flash-Chat E2P10\n- org=meituan-longcat · release_type=model weights · artifact=LongCat-Video (text-to-video) · traction=1,708 downloads / 569 likes · evidence_url=https://huggingface.co/meituan-longcat/LongCat-Video E1P18\n- org=meituan-longcat · release_type=model weights · artifact=LongCat-Image / Image-Edit / Edit-Turbo (image gen+edit) · traction=13,738 / 20,127 / 23,329 downloads · evidence_url=https://huggingface.co/meituan-longcat/LongCat-Image E4E7E13P20\n- org=meituan-longcat · release_type=model weights · artifact=LongCat-Flash-Omni (560B any-to-any) · traction=198 downloads / 116 likes · evidence_url=https://huggingface.co/meituan-longcat/LongCat-Flash-Omni E10P16\n- org=meituan-longcat · release_type=model weights · artifact=LongCat-Next (74B any-to-any) · traction=287 downloads / 212 likes · evidence_url=https://huggingface.co/meituan-longcat/LongCat-Next E5P28\n- org=meituan-longcat · release_type=model weights · artifact=LongCat-Flash-Lite / -Sparse (69B non-thinking MoE, LSA 1M-context) · traction=7,333 / 609 downloads · evidence_url=https://huggingface.co/meituan-longcat/LongCat-Flash-Lite-Sparse E6E11P2\n- org=meituan-longcat · release_type=model weights · artifact=LongCat-Flash-Prover (560B, Lean4 formal reasoning) · traction=195 downloads / 35 likes · evidence_url=https://huggingface.co/meituan-longcat/LongCat-Flash-Prover E19P27\n- org=meituan-longcat · release_type=model weights · artifact=LongCat-AudioDiT 1B/3.5B (voice cloning) + LongCat-Audio-Codec · traction=1,395 / 1,670 downloads · evidence_url=https://huggingface.co/meituan-longcat/LongCat-AudioDiT-3.5B E12E15E17W4\n- org=meituan-longcat · release_type=code/repo · artifact=LongCat-DeepResearch harness (MIT, Python) · traction=8 stars at repo creation · evidence_url=https://github.com/meituan-longcat/LongCat-DeepResearch P1E16\n\nTalking\n- org=meituan-longcat · theme=domestic-compute sovereignty · public_framing=LongCat-2.0 as the industry's first trillion-parameter model fully trained+served on ~50k domestic cards, with domestic-chip-optimized inference code released · evidence_url=https://www.36kr.com/p/3884214101324041 W1\n- org=meituan-longcat · theme=agentic AI R&D evaluation · public_framing="Beyond Final Scores" — 7 frontier models evaluated on 36 AutoLab tasks, 756 trajectories, ~$100k inference; only 3 of 252 best solutions judged novel · evidence_url=https://www.laura-martel.com/blog/longcat-autoresearcheval-engineering-optimizers W2\n- org=meituan-longcat · theme=LongCat-2.0 architecture/deployment · public_framing=LSA sparse attention, N-gram Embedding, 6D parallel training on domestic compute, prefill-decode disaggregated deployment; SGLang (GPU) + SGLang-FluentLLM (NPU) entry points · evidence_url=https://go.tabbit.ai/model/longcat-2-0/reviews/official-technical-blog-architecture-training-inference-deployment W6\n- org=meituan-longcat · theme=audio/voice cloning · public_framing=LongCat-AudioDiT beats prior zero-shot voice cloning limits via single-stage training on ASR transcription data · evidence_url=https://www.xix.ai/ainews/meituan-opensource-audio-model-sets-new-voice-cloning-benchmark.html W4\n- org=meituan-longcat · theme=GTM/platform · public_framing=LongCat-2.0 launch with API platform (longcat.chat) and token giveaway/promo pricing · evidence_url=http://www.lpqz.cn/news/3745 W3\n- org=meituan-longcat · theme=community attention · traction=HN: LongCat-Flash-Chat 4 points/0 comments; LongCat-Flash-Thinking 2 points/1 comment · evidence_url=https://github.com/meituan-longcat/LongCat-Flash-Chat E21E23\n\n## Shipping\n\nLongCat's shipping cadence is dense and multimodal. The flagship LongCat-2.0 (1.6T total / ~48B activated MoE) shipped on HF with MIT license, built and deployed entirely on AI ASIC superpods P5P6E3. Around it: LongCat-Flash-Chat (560B) E2P10, LongCat-Flash-Thinking (560B reasoning) and its -2601 and -ZigZag variants E8E9E20P9, LongCat-Flash-Omni (560B any-to-any) E10P16, LongCat-Next (74B any-to-any) E5P28, the 69B non-thinking LongCat-Flash-Lite/-Sparse with 1M-token LSA context E6E11P2, and LongCat-Flash-Prover for Lean4 formal reasoning E19P27.\n\nGeneration-side releases span video (LongCat-Video, text-to-video) E1P18, image (LongCat-Image, Image-Edit, Edit-Turbo, Image-Dev) E4E7E13E14P20, and audio (LongCat-Audio-Codec tokenizer; LongCat-AudioDiT 1B/3.5B) E17E12E15W4P11. Systems releases include SGLang-FluentLLM (inference engine + kernels, Apache-2.0) P24, omni-flow (workflow orchestration + distributed KV-cache sharing for multimodal inference) and its SGLang-integration repo omni-flow-sglang P3P4, plus LongCat-DeepResearch, an open-source deep-research harness paired with LongCat-2.0 P1E16.\n\n## Research themes\n\n- Sparse attention & long context. LongCat-Flash-Lite-Sparse replaces dense MLA with LongCat Sparse Attention (LSA), extending DeepSeek Sparse Attention (DSA) with indexing mechanisms and native 1M-token context P2; LSA is also central to LongCat-2.0's architecture W6.\n- Reasoning & formal methods. A dedicated reasoning line (LongCat-Flash-Thinking, -2601, ZigZag) P9E9E20 plus LongCat-Flash-Prover for "Native Formal Reasoning in Lean4" via agentic tool-integrated reasoning P27. R-HORIZON studies LRM breadth/depth degradation over long reasoning horizons P14E45.\n- Multimodal generation & omni models. Video, image, audio codec/DiT, and any-to-any omni models E1E4E10E12P11P16P18P20; UNO-Bench probes the "compositional law" between uni-modal and omni-modal capabilities P15E39.\n- Agents & deep research. LongCat-DeepResearch (ResearchSpec, independent section research, targeted report revisions) P1; DPT-Agent (dual-process human-AI collaboration, ACL 2025) P19; MineExplorer (open-world MLLM exploration in Minecraft) P7E44; agentic AI-R&D evaluation via AutoLab W2.\n- Evaluation as a first-class output. VitaBench (ICLR 2026 tool-use, cited by Qwen3.5 and Seed2.0) P13, AMO-Bench (math competitions; Gemini 3 Pro and Qwen3-Max-Thinking SOTA updates) P17, Meeseeks (multi-turn instruction-following self-correction) P8, plus SOP-Maze, PIE_bench, WBench, General365, LARYBench, VitaBench-2.0 P12P21E32E38E34E40.\n- Model–system co-design. SGLang-FluentLLM refactors speculative decoding, fuses Target+Verify+Draft into one CUDA graph, adds layer-wise KV-cache transfer and a decode radix-tree cache, and open-sources kernels (FlashMLA, DeepGEMM, FlashInfer) P24; omni-flow builds a three-layer (control/data/compute) abstraction over SGLang for multimodal KV-cache sharing P3P4.\n\n## Hiring & scaling\n\nHiring evidence is thin: the pack contains a single item — the "LongCat 2026" large-model internship program aimed at full-time master's/PhD students worldwide, covering base-model algorithms, infrastructure architecture, agent applications, and visual generation W5. This is consistent with the lab's research breadth (models, systems, agents, multimodal) but no full-time roles, headcount, or location/hub data are cited. The gap itself is notable: the evidence proves activity across 20+ repos and 15+ model releases but does not prove hiring scale P1P18E1E3.\n\n## Category implications\n\n- Strategy / compute sovereignty. LongCat-2.0's full training+inference on domestic AI-ASIC superpods and the release of NPU inference code (SGLang-FluentLLM/tree/npu) imply a deliberate hedge against GPU supply constraints and a bid for the domestic-compute ecosystem P5W1W3W6. This reframes LongCat not just as a model lab but as a reference implementation for training/serving frontier models on domestic accelerators.\n- Infrastructure. The Feb-2026 fork cluster (FlashMLA, DeepGEMM, DeepEP, flashinfer, mscclpp, fast-hadamard-transform) plus SGLang-FluentLLM and omni-flow signal heavy investment in custom inference kernels, speculative decoding, KV-cache management, and multimodal serving — infrastructure that unlocks both cost-efficient long-context and multimodal deployment P24P3P4E49E52E56E57E58E54.\n- Product / GTM. A launch-time API platform (longcat.chat) with token promos, combined with ModelScope distribution, indicates a consumer/developer GTM motion beyond raw open weights W3. Deep-research (LongCat-DeepResearch) and voice cloning (LongCat-AudioDiT) point to agentic and speech product surfaces P1W4.\n- Research. The lab is building a self-reinforcing eval moat — multiple benchmarks (VitaBench, AMO-Bench, UNO-Bench, Meeseeks, R-HORIZON) that third-party labs (Qwen, Seed, Gemini, GLM, Kimi) adopt or appear on, which grants LongCat agenda-setting influence in agent, reasoning, and omni evaluation P13P17P8P14P15.\n- Hiring. The internship program's stated areas map directly to the lab's shipped surface (base models, infrastructure, agents, visual generation), implying near-term demand for those specializations W5.\n\n## Traction highlights\n\nGitHub stars (repo metadata snapshots): LongCat-Video 4,273 stars / 669 forks P18; LongCat-Flash-Chat 1,339 stars / 67 forks P10; LongCat-Image 695 stars / 60 forks P20; LongCat-Flash-Omni 492 stars / 33 forks P16; LongCat-Next 438 stars / 23 forks P28; LongCat-Audio-Codec 301 stars / 23 forks P11; LongCat-Flash-Thinking 285 stars / 29 forks P9; LongCat-2.0 41 stars / 2 forks (page snapshot, early post-release) P6. Hugging Face downloads: LongCat-Image-Edit-Turbo 23,329; LongCat-Flash-Chat 23,871; LongCat-Image-Edit 20,127; LongCat-Image 13,738 E13E2E7E4. Third-party adoption: VitaBench cited by Qwen3.5 and Seed2.0 P13.\n\n## Sources\n\n- P1 meituan-longcat/LongCat-DeepResearch — repo metadata (harness, MIT)\n- P2 LongCat-Flash-Lite-Sparse — model card (69B MoE, LSA, 1M context)\n- P3 meituan-longcat/omni-flow — repo metadata (multimodal inference orchestration)\n- P4 meituan-longcat/omni-flow-sglang — SGLang modifications repo\n- P5 LongCat-2.0 — HF model card (1.6T/48B, AI ASIC superpods)\n- P6 meituan-longcat/LongCat-2.0 — repo metadata\n- P7 meituan-longcat/MineExplorer — repo metadata (Minecraft MLLM benchmark)\n- P8 meituan-longcat/Meeseeks — repo metadata (instruction-following benchmark)\n- P9 meituan-longcat/LongCat-Flash-Thinking — repo metadata (560B LRM)\n- P10 meituan-longcat/LongCat-Flash-Chat — repo metadata (560B)\n- P11 meituan-longcat/LongCat-Audio-Codec — repo metadata\n- P12 meituan-longcat/SOP-Maze — repo metadata\n- P13 meituan-longcat/vitabench — repo metadata (ICLR 2026)\n- P14 meituan-longcat/R-HORIZON — repo metadata\n- P15 meituan-longcat/UNO-Bench — repo metadata\n- P16 meituan-longcat/LongCat-Flash-Omni — repo metadata\n- P17 meituan-longcat/AMO-Bench — repo metadata\n- P18 meituan-longcat/LongCat-Video — repo metadata\n- P19 meituan-longcat/DPT-Agent — repo metadata\n- P20 meituan-longcat/LongCat-Image — repo metadata\n- P21 meituan-longcat/PIE_bench — repo metadata\n- P24 meituan-longcat/SGLang-FluentLLM — repo metadata (engine + kernels)\n- P27 meituan-longcat/LongCat-Flash-Prover — repo metadata (Lean4)\n- P28 meituan-longcat/LongCat-Next — repo metadata\n- E1–E58 LongCat release/repo/fork events (HF downloads/likes, repo stars, fork parents)\n- W1 36kr — LongCat-2.0 domestic-compute open-source report\n- W2 Laura Martel — LongCat AutoLab agentic R&D evaluation\n- W3 lpqz.cn — LongCat-2.0 open-source + inference code + API platform\n- W4 xix.ai — LongCat-AudioDiT voice cloning benchmark\n- W5 xix.ai — LongCat 2026 internship program\n- W6 Tabbit — LongCat-2.0 official technical blog summary","cites":["P1","P2","P3","P4","P5","P6","P7","P8","P9","P10","P11","P12","P13","P14","P15","P16","P17","P18","P19","P20","P21","P24","P27","P28","E1","E2","E3","E4","E5","E6","E7","E8","E9","E10","E11","E12","E13","E14","E15","E16","E17","E19","E20","E21","E23","E24","E32","E34","E36","E38","E39","E40","E45","E49","E52","E54","E56","E57","E58","W1","W2","W3","W4","W5","W6"]}