← Analysis archivereport.jsonevidence.jsonDossier →
Meituan (LongCat)Neolabgenerated Aug 7, 2026 · 3w

Meituan (LongCat) analysis

62 cited sources57 signals on file

Thesis

Meituan's LongCat lab has executed one of the fastest documented progressions from modest open-source releases to a 1.6-trillion-parameter, near-frontier agentic coding model — trained entirely on Chinese AI ASIC superpods — in roughly ten months P4W1W2. The lab pursues an explicit model–system co-design philosophy, pairing each model family with custom inference kernels and a growing benchmark portfolio that doubles as competitive intelligence and academic credibility P23P12P13P16. The LongCat-2.0 launch, combined with the concurrent release of Omni-Flow for multimodal inference orchestration, signals a lab transitioning from research exploration to production-grade infrastructure capable of serving Meituan's core delivery and local-services platform at scale P2P4W4.

Signal desks

Hiring

No cited evidence in this pack. The available evidence contains no job postings, role descriptions, or team-expansion announcements for LongCat.

Forks

  • DeepSeek-ai/FlashMLA — Forked with custom feature/swapAB and feature/ckv_fp8_per_token branches for FP8 KVCache + FP8 compute optimizations, tied to the SnapMLA paper E51P23.
  • DeepSeek-ai/DeepGEMM — Forked with feature/swap_ab branch for SwapAB Offset + PDL optimizations E55P23.
  • DeepSeek-ai/DeepEP — Forked for expert-parallel communication tuning E56.
  • flashinfer-ai/flashinfer — Forked with feature/longcat_main branch for communication–computation fused kernel optimizations E57P23.
  • Dao-AILab/fast-hadamard-transform — Forked, likely related to quantization or attention kernel work E53.
  • microsoft/mscclpp — Forked for custom collective-communication primitives E48.

All forks cluster in early February 2026, coinciding with the SGLang-FluentLLM inference-engine release, suggesting a concentrated push to harden the inference stack for the upcoming LongCat-2.0 deployment E39P23. The fork pattern reveals heavy dependency on the DeepSeek open-source inference ecosystem (FlashMLA, DeepGEMM, DeepEP) adapted for domestic AI ASICs W2P4.

Releases

  • LongCat-Flash-Chat (560B MoE, text-generation) — Aug 2025. 55K+ HuggingFace downloads, 537 likes. Flagship debut release E2P9.
  • LongCat-Flash-Thinking (560B LRM) — Sep 2025. Reasoning model with 140 downloads, 148 likes E8P8.
  • LongCat-Flash-Omni (560B any-to-any) — Oct 2025. Multimodal any-to-any model E11P15.
  • LongCat-Video — Oct 2025. Text-to-video generation, 1.8K downloads, 546 likes E1P17.
  • LongCat-Audio-Codec — Oct 2025. Audio tokenizer/detokenizer for speech LLMs E16P10.
  • LongCat-Image (text-to-image) — Dec 2025. Apache-2.0 license, 13.3K downloads E4P19.
  • LongCat-Image-Edit (image-to-image) — Dec 2025. 20.6K downloads, highest-download image variant E7P19.
  • LongCat-Flash-Thinking-2601 — Jan 2026. Updated reasoning model, 3K downloads E9P21.
  • LongCat-Flash-Thinking-ZigZag — Jan 2026. Experimental reasoning variant E19P22.
  • LongCat-Flash-Lite (69B, text-generation) — Jan 2026. Dense MLA model, 17.2K downloads E6.
  • LongCat-Image-Edit-Turbo — Feb 2026. 22.5K downloads, fastest image-edit variant E13P19.
  • LongCat-Flash-Prover (560B, Lean4 formal reasoning) — Mar 2026. Agentic tool-integrated formal reasoning E17P26.
  • LongCat-Next (74B, any-to-any) — Mar 2026. Native multimodal model E5P27.
  • LongCat-AudioDiT-1B / 3.5B — Mar 2026. Audio diffusion transformer models E12E15.
  • LongCat-2.0 (1.6T MoE, ~48B activated, text-generation) — Jun/Jul 2026. MIT license. Trained entirely on AI ASIC superpods. 2.6K downloads, 232 likes in first weeks. Full weights uploaded as 194 safetensors shards (~3.6TB total) E3P4W3.
  • LongCat-Flash-Lite-Sparse (69B total, ~3B activated, 1M-token context) — Jul/Aug 2026. Replaces dense MLA with LongCat Sparse Attention (LSA), built on Flash-Lite E10P1.
  • Omni-Flow — Jun 2026. Workflow orchestration and distributed KV cache sharing framework for multimodal inference P2E22.
  • SGLang-FluentLLM — Feb 2026. Custom SGLang-based inference engine with speculative decoding (Eagle/MTP/PLD), layer-wise KVCache transfer, Decode Radix Tree Cache P23E39.

Talking

  • VentureBeat (Carl Franzen, Jun 2026): Coverage of LongCat-2.0 as the unmasked model behind "Owl Alpha" — the anonymous model leading OpenRouter charts for two months. Highlights training entirely on Chinese chips and the 1M-token context window under MIT license. Notes weights "coming soon" at time of writing W1.
  • Meituan Tech Blog (Jul 2026): Official open-source announcement emphasizing LongCat-2.0 as the industry's first trillion-parameter model inferenced on a 50K-card domestic compute cluster. Frames the release as unlocking value from existing domestic chip stock. References SGLang PR #30042 for NPU integration W2.
  • eesel AI (Aug 2026): Analyst deep-dive comparing LongCat-2.0 to Kimi K2.7 Code and GPT-5.1-Codex-Max. Notes Meituan retired six Flash models a month before the 2.0 launch to consolidate the product line. Frames the model as agentic-coding-focused rather than chat W4.
  • Polymath707 / Substack (Jul 2026): Background on LongCat's origin via Meituan's mid-2023 acquisition of Light Years Beyond (Guangnian Zhi Wai), the AI startup founded by Meituan co-founder Wang Huiwen with $50M of personal capital. Positions the lab within the wave of Chinese frontier labs competing with OpenAI and Anthropic W5.

Shipping

LongCat has shipped 15+ distinct model artifacts across five modalities (text, image, video, audio, code) since August 2025, with a clear cadence acceleration in late 2025 and early 2026 . The most downloaded models are LongCat-Flash-Chat (55K+), LongCat-Image-Edit-Turbo (22.5K), and LongCat-Image-Edit (20.6K), suggesting strong practical pull for image-editing capabilities E2E13E7. Beyond models, the lab shipped SGLang-FluentLLM (a custom inference engine), Omni-Flow (multimodal orchestration), and LongCat-Next-inference (a multimodal inference testing framework), indicating serious investment in the deployment layer P23P2P28. LongCat-2.0's full-weight release via 194 safetensors shards makes it one of the largest openly distributed models, though the evidence shows weights were uploaded only after an initial code/config-only release W3W1. Meituan retired six Flash-branded models prior to the 2.0 launch, signaling a brand consolidation strategy W4.

Research themes

1. Extreme-scale MoE with sparse attention — LongCat-2.0 (1.6T/48B activated) and LongCat-Flash-Lite-Sparse (69B/3B activated) push MoE efficiency further with LongCat Sparse Attention (LSA), extending DeepSeek Sparse Attention with streaming-aware indexing and hardware-efficient sparse patterns for up to 1M-token context P4P1.

2. Model–system co-design — The lab explicitly frames its work around this principle. SGLang-FluentLLM refactors speculative decoding (Eagle, MTP, PLD), combines target/verify/draft into single CUDA graphs, and implements layer-wise KVCache transfer with overlap scheduling P23. Omni-Flow extends co-design into the multimodal domain with a three-layer abstraction (Control/Data/Compute Flow) built on SGLang P2P3.

3. Full-modality coverage — LongCat has released distinct models or frameworks for text generation, reasoning, image generation/editing, video generation (including avatar models), audio codec/tokenization, speech synthesis (AudioDiT), and any-to-any native multimodality (Flash-Omni, Next) E1E4E5E11E12E16P10P17P19P27.

4. Formal reasoning and proof — LongCat-Flash-Prover targets Native Formal Reasoning in Lean4 for mathematics formalization and theorem proving through agentic tool-integrated reasoning P26E17.

5. Agent evaluation at scale — A distinct research thread builds benchmarks for agentic capabilities: VitaBench (tool-use, ICLR 2026, cited by Qwen3.5 and Seed2.0), MineExplorer (open-world MLLM exploration in Minecraft), R-HORIZON (long-horizon reasoning breadth/depth, ICLR 2026), SOP-Maze (business SOP execution), DPT-Agent (dual-process human-AI collaboration, ACL 2025), and WBench (interactive video world models) P12P6P13P11P18E29.

6. Math and general reasoning benchmarks — AMO-Bench (high-school math competitions, multiple SOTA claims tracked), UNO-Bench (omni-model compositional law), General365 (general reasoning under high difficulty/diversity), and LARYBench P16P14E41E37.

7. Domestic chip deployment — LongCat-2.0 was both trained and inferenced entirely on Chinese AI ASIC superpods, with specific optimizations for memory-bandwidth-constrained domestic chips. The Meituan tech blog describes deep co-optimization across model architecture, chip adaptation, and deployment strategy P4W2.

Hiring & scaling

No hiring data (job postings, role descriptions, headcount figures) is present in this evidence pack. However, several indirect scaling signals are visible:

  • The lab originated from Meituan's mid-2023 acquisition of Light Years Beyond, founded by Meituan co-founder Wang Huiwen with $50M in personal capital and investments from CEO Wang Xing W5.
  • The breadth of simultaneous research streams — from video generation (4,273 GitHub stars) to formal proof to agent benchmarks — implies a multi-team organizational structure P17P26P12.
  • LongCat-2.0's 50K-card domestic compute cluster deployment and the Omni-Flow/Omni-Flow-SGLang infrastructure release suggest significant DevOps and systems-engineering headcount beyond pure research W2P2P3.
  • The pace of model releases (15+ in ~12 months) and six-model retirement before LongCat-2.0 points to an active product-management and release-engineering function W4.

Category implications

Infrastructure / silicon strategy: LongCat-2.0 being trained and inferenced entirely on Chinese AI ASIC superpods is a category-significant signal P4W2. The lab's deep forking of DeepSeek's inference stack (FlashMLA, DeepGEMM, DeepEP) with custom optimizations for domestic hardware — combined with the Omni-Flow/SGLang pipeline — suggests LongCat is building a vertically integrated inference stack that reduces dependency on NVIDIA CUDA E51E55E56P23. The Meituan tech blog's explicit framing around "unlocking value from existing domestic chip stock" and the reference to NPU SGLang integration (PR #30042) indicate this is a strategic priority, not just a research curiosity W2.

Product / GTM strategy: The lab has pivoted from general-purpose chat (LongCat-Flash-Chat) toward agentic coding as the primary product surface W4W1. LongCat-2.0's positioning against Kimi K2.7 Code and GPT-5.1-Codex-Max, combined with its stealth deployment as "Owl Alpha" on OpenRouter for two months before the public unveiling, reveals a launch strategy that validates product-market fit anonymously before branding W1W4. The MIT license on the flagship 1.6T model signals a commercially permissive open-weight strategy, likely aimed at enterprise adoption P4W1.

Benchmarks as competitive moat: LongCat has released at least 11 distinct benchmarks spanning tool-use (VitaBench, ICLR 2026), math (AMO-Bench), reasoning horizons (R-HORIZON, ICLR 2026), omni-model evaluation (UNO-Bench), instruction-following (Meeseeks), business SOPs (SOP-Maze), agent exploration (MineExplorer), general reasoning (General365), video world models (WBench), and multilingual QA (PIE_bench) P12P16P13P14P7P11P6E41E29P20. Several have been adopted by competitors (Qwen, ByteDance Seed) for their own model evaluations, and VitaBench was cited by Qwen3.5 and Seed2.0 P12. This creates an asymmetric intelligence advantage: LongCat designs evaluation frameworks that competitors then use, giving LongCat insight into how rival models perform against criteria LongCat defined.

Multimodal pipeline lock-in: Omni-Flow's three-layer architecture (Control/Data/Compute Flow) built on SGLang — with distributed KV cache sharing, multimodal prefix matching, and unified diffusion/LLM forward paths — is positioned as infrastructure that could become a dependency for any team building multimodal inference pipelines P2P3. The companion omni-flow-sglang repo integrates these modifications directly into SGLang, suggesting an upstream-contribution strategy P3.

Research-to-production compression: The ~10-month timeline from the first Flash-Chat release (Aug 2025) to LongCat-2.0 (Jun 2026) — spanning 560B to 1.6T parameters — demonstrates unusually fast iteration for a lab of this profile E2E3. Combined with the fact that the lab was formed via acquisition only in mid-2023, the velocity implies either significant pre-existing infrastructure or aggressive talent concentration W5.

Traction highlights

  • LongCat-Video: 4,273 GitHub stars and 669 forks, the highest-engagement repo in the LongCat portfolio P17.
  • LongCat-Flash-Chat: 1,339 GitHub stars, 67 forks, and 55K+ HuggingFace downloads — the most downloaded text model E2P9.
  • LongCat-Image-Edit-Turbo: 22.5K downloads, the highest-download image model variant E13.
  • LongCat-2.0: 41 GitHub stars at launch (growing to 513+), 232 HuggingFace likes in the first weeks E3E20.
  • LongCat-Audio-Codec: 301 stars, 23 forks P10.
  • LongCat-Flash-Omni: 492 stars, 33 forks P15.
  • LongCat-Next: 438 stars, 23 forks P27.
  • LongCat-Image: 695 stars, 60 forks P19.
  • VitaBench: ICLR 2026 acceptance, cited by Qwen3.5 and ByteDance Seed2.0 P12.
  • R-HORIZON: ICLR 2026 acceptance P13.
  • DPT-Agent: ACL 2025 main conference acceptance P18.
  • Owl Alpha / OpenRouter: LongCat-2.0 operated anonymously as the top model on OpenRouter for two months before the official unveiling W1.