A few hours ago Meituan, the Chinese delivery app and superapp company, publicly released LongCat-2.0 on GitHub, Hugging Face and its own platform. Meituan disclosed that the model was the engine behind the anonymous "Owl Alpha" instance that dominated developer charts on OpenRouter over the past two months.
Model and license
LongCat-2.0 is a 1.6-trillion-parameter Mixture-of-Experts (MoE) architecture with a native one-million-token context window. Meituan released the model under the MIT license, allowing broad commercial use and modification without copyleft obligations.
Training and hardware
According to Meituan, the model was trained entirely on a cluster of over 50,000 domestically produced application-specific integrated circuits (ASICs). This demonstrates that near-frontier large models can be scaled without relying on U.S. Nvidia GPUs.
Pricing and commercial access
Meituan offers commercial access via both a traditional pay-as-you-go API and structured "Token Pack" packages. Standard API pricing for uncached requests is $0.75 per million tokens for input and $2.95 per million tokens for output, while a limited-time promotional price reduces these to $0.30 (input) and $1.20 (output) per million tokens.
A distinctive commercial feature is that context-cache hits are processed free of charge: only cache-miss inputs and final token generations consume Token Pack quotas. Token Packs are fixed-volume allocations valid for 30 days and sold in limited flash sales four times daily at 10:00, 16:00, 21:00 and 23:00 Beijing Time on a first-come, first-served basis.
In comparative market tables, the LongCat-2.0 promotional rates sit on the lower-cost end among top-performing models.
Benchmarks and specialization
LongCat-2.0 is explicitly optimized for multi-step engineering tasks, tool integration and automated repository manipulation. On standard tests it scores 59.5 on SWE-bench Pro, versus GPT-5.5's 58.6. Other reported results include 70.8 on Terminal-Bench 2.1, 77.3 on SWE-bench Multilingual and 73.2 on the FORTE corporate workflow simulator.
Technology: sustaining a 1M-token sparse context
The model aggressively optimizes MoE sparsity: while total parameters reach 1.6 trillion, active computation averages about 48 billion parameters per token (ranging roughly 33–56 billion depending on query complexity). The design uses a "Zero-Compute Experts" approach so routine operations are routed through lighter subnetworks to avoid idle compute overhead.
To keep a functional one-million-token window without catastrophic hardware bottlenecks, Meituan developed LongCat Sparse Attention (LSA). LSA addresses quadratic scoring costs and memory fragmentation via three orthogonal mechanisms:
- Streaming-aware Indexing (SI): blends hardware-aligned contiguous reads with dynamic random selection, converting fragmented accesses into predictable sequential blocks for improved HBM utilization and effective bandwidth.
- Cross-Layer Indexing (CLI): leverages the empirical stability of attention saliency across adjacent hidden layers so a single indexing pass can guide multiple layers, reinforced during training by cross-layer distillation.
- Hierarchical Indexing (HI): applies a coarse-to-fine two-stage scoring layout — fast block-level recall filters candidates, then fine-grained token selection runs on the reduced set.
Meituan also integrated an N-gram Embedding module from its lighter models. A 5-gram token combination framework adds 135 billion parameters in sparse dimensions orthogonal to the MoE experts, expanding the core embedding space roughly 100-fold to better capture dense local token relations and reduce memory I/O during large-batch inference.
Post-training: MOPD and specialized experts
LongCat-2.0’s post-training stage uses Multi-Teacher Optimization via Mixture of Specialized Experts (MOPD). Instead of collapsing feedback into a single reward model, MOPD divides optimization into three independent expert clusters:
- Agent Experts: tuned for structural execution, precise tool invocation, multi-turn API parsing and self-correcting execution loops.
- Reasoning Experts: optimized for multi-hop logic, chain-of-thought, mathematics and higher-level STEM problem solving.
- Interaction Experts: focused on instruction-following, factual grounding and safety guardrails.
These vectors are trained separately and during runtime a dynamic gate-routing mechanism fuses them so the final system coordinates deep reasoning, reliable tool execution and safe user interaction simultaneously.
Market and geopolitical context
The release is notable against the backdrop of U.S. pressure on Western labs to limit access to certain models. The article notes that, following a U.S. governmental request, OpenAI restricted access to GPT-5.6, and that Anthropic was ordered to limit access to Claude Fable 5 / Mythos 5. Observers argue that locking down Western closed-source models and rising API costs create room for affordable, high-performance alternatives provided by Chinese open-source releases like Meituan's LongCat-2.0.
During its unbranded run on OpenRouter as Owl Alpha, the model generated about 10.1 trillion monthly tokens — roughly 559 billion tokens per day — a 242% month-over-month surge that placed it in the platform’s global top three. By the time Meituan claimed the architecture, the model had reached top positions across several deployment environments.
Meituan’s strategic shift
Meituan was founded in March 2010 by Wang Xing. After merging with Dianping in 2015, the Beijing-based company expanded into a dominant superapp in China. Meituan reports over 770 million annual transacting users and supports more than 14.5 million merchants. Facing margin pressure and intense domestic competition, the firm redirected substantial resources into AI and domestic chip development, positioning itself toward global AI infrastructure.
Earlier steps included the late-2025 release of LongCat-Flash (560 billion parameters) and LongCat-Flash-Thinking, both open-sourced under enterprise-friendly licenses.
Enterprise implications
An open-weight, MIT-licensed model with a one-million-token context window enables enterprises to run large-scale autonomous software workflows locally and avoid recurring third-party API costs. Use cases Meituan highlights include repository-level migration and refactoring: an organization can feed entire codebases plus modern SDK documentation into the model’s context window for dependency mapping, structural updates, compilation and automated bug-catching within local sandboxes prior to generating a pull request.
The MOPD gate-routing mechanism offers compliance advantages by isolating deep reasoning passes from interaction safeguards, letting sensitive sectors (finance, healthcare) run complex logic while limiting hallucination risks. Combined with free cache-hit processing, this can materially lower operational costs for repeated, large-context agentic workflows.
Conclusion
LongCat-2.0’s public release combines a permissive MIT license, a native one-million-token context window and domestic ASIC-based training to signal a potential structural shift in AI infrastructure. Its pricing model, cache-based economics and MOPD specialization make it specifically aimed at autonomous software engineering, and the release marks another step in Meituan’s transformation from a regional superapp into an infrastructure player in the global AI landscape.
