LongCat-2.0: How Meituan Trained a 1.6T-Parameter Coding Model Without a Single Nvidia GPU

Meituan — best known in China as a food delivery and super-app company — quietly ran one of the most technically ambitious open-weight model releases of 2026. LongCat-2.0 is a 1.6-trillion-parameter Mixture-of-Experts (MoE) model built for agentic coding tasks, released under the MIT license on June 30, 2026, with full weights available on Hugging Face as of July 5. What makes it worth examining is not just the scale, but two specific engineering choices: a novel sparse attention mechanism designed for million-token contexts, and a training run conducted entirely on domestic Chinese AI ASICs — no Nvidia hardware involved.

What LongCat-2.0 Is Built For

LongCat-2.0 is not a general-purpose chat model. It is explicitly designed for autonomous software engineering: repository-level code edits, multi-step tool use, and long-horizon agent workflows. Meituan measured it through Claude Code sandboxes and the Hermes agent harness, and the benchmarks reflect that focus.

On SWE-bench Pro, LongCat-2.0 scores 59.5, narrowly ahead of GPT-5.5 at 58.6. On Terminal-Bench 2.1, it scores 70.8, comparable to Gemini 3.1 Pro (70.7). It trails Claude Opus 4.8 (69.2 on SWE-bench Pro, 78.9 on Terminal-Bench) on coding benchmarks, and falls behind on broader general-agent tasks like BrowseComp. The model's strength is narrow and deliberate: it is competitive with closed-source frontier models specifically on software engineering, at a fraction of the API cost.

Before the official reveal, LongCat-2.0 operated anonymously on OpenRouter as "Owl Alpha," where it accumulated roughly 10.1 trillion monthly tokens — a 242% month-over-month increase — and reached the top of the Hermes Agent leaderboard. Developers were using it without knowing who built it.

The Sparse Attention Problem at 1M Tokens

Running a 1-million-token context window efficiently is genuinely hard. Standard attention scales quadratically with sequence length, and even sparse attention mechanisms can create memory fragmentation and unpredictable bandwidth usage at this scale.

Meituan's answer is LongCat Sparse Attention (LSA), which extends DeepSeek Sparse Attention with three orthogonal improvements:

Streaming-aware Indexing (SI) restructures how tokens are selected for attention. Instead of fragmented random reads from GPU high-bandwidth memory (HBM), SI converts token selection into sequential, hardware-aligned memory reads. The result is coalesced HBM access — the memory controller can prefetch predictably rather than chasing scattered addresses.

Cross-Layer Indexing (CLI) exploits a property of transformer attention: which tokens matter tends to be stable across adjacent layers. CLI computes one indexing pass and reuses it across multiple consecutive layers during inference. This amortizes the cost of figuring out which tokens to attend to, and the stability assumption is reinforced during training via cross-layer distillation.

Hierarchical Indexing (HI) applies a two-stage scoring approach. A fast, approximate block-level pass filters the token population down to candidates, and then fine-grained token selection runs only on that smaller set. This avoids scoring every token at full precision.

LSA also extends to Multi-Token Prediction (MTP) for speculative decoding, which can further improve throughput at inference time.

On top of LSA, the model adds 135 billion N-gram Embedding parameters (using 5-gram token combinations). This expands the embedding space roughly 100-fold without adding to the MoE expert count, and it reduces memory I/O during large-batch decoding by shifting parameter weight away from the expert routing path.

Training on 50,000 Chinese ASICs

The hardware story is the part that has attracted the most attention. Meituan trained LongCat-2.0 on a cluster of over 50,000 domestic Chinese AI ASICs — reportedly Huawei Ascend-class accelerators — without any Nvidia GPUs. The training run covered 35+ trillion tokens with no rollbacks or irrecoverable loss spikes.

This is not trivial. Frontier-scale training runs on non-standard hardware typically encounter instability from non-deterministic floating-point operations, fault recovery gaps, and memory parallelism issues that are well-solved on Nvidia's CUDA stack but require custom engineering elsewhere. Meituan addressed this through deterministic operators across embedding, flash attention, LSA, and MoE layers; automated link isolation and rejoin after hardware stress tests; and a 6D parallelism scheme that includes a custom approach called EMBP for the N-gram Embedding module.

The cluster is organized into superpods of up to 48 machines each, with all-to-all communication inside each pod and RoCE (RDMA over Converged Ethernet) between pods. Meituan reports approximately 30% pretraining throughput gain from this topology.

The significance here is structural. U.S. export controls have restricted access to Nvidia's highest-end chips for Chinese companies. If a 1.6T-parameter model can be trained to near-frontier performance on alternative silicon, it suggests that export restrictions on GPU hardware may have less long-term effect on Chinese AI development than initially assumed.

Post-Training: Three Expert Clusters

LongCat-2.0 uses a post-training framework called MOPD (Multi-Teacher Optimization via Mixture of Specialized Experts). Rather than a single reward model, MOPD trains three independent expert clusters:

• Agent Experts handle tool invocation, multi-turn API parameter parsing, and self-correcting execution loops.
• Reasoning Experts focus on multi-hop logic, mathematics, and STEM problem-solving, with adaptive compute allocation based on problem difficulty.
• Interaction Experts handle instruction following, hallucination suppression, and safety constraints.

A dynamic gate-routing mechanism combines these at runtime. The separation is meant to prevent the common failure mode where safety tuning degrades reasoning performance, or where reasoning optimization causes the model to ignore instructions.

Access and Licensing

The model is available on Hugging Face under the MIT license, with BF16/F32 safetensors and an FP8 variant. GPU inference runs via SGLang (16x H20 recommended); NPU inference is supported through SGLang-FluentLLM. Self-hosting at this scale requires significant infrastructure — multiple H20 nodes — so it is not a laptop experiment.

Via API, Meituan offers standard pay-as-you-go pricing ($0.75/$2.95 per million tokens in/out) with a promotional rate of $0.30/$1.20. A notable feature: context cache hits are processed free of charge, which meaningfully changes the economics for agentic workflows that repeatedly reference the same large codebase.

What This Demonstrates

LongCat-2.0 is worth paying attention to for two reasons that go beyond benchmark numbers. First, it shows that sparse attention for million-token contexts is a solvable engineering problem — LSA's three-component approach is a concrete example of how to make long-context inference practical at scale. Second, the training infrastructure story is a data point about the state of non-Nvidia AI hardware: a 35-trillion-token run on 50,000 ASICs, completed without instability, is evidence that the hardware gap is narrowing.

For developers, the MIT license and open weights mean LongCat-2.0 can be integrated into commercial products without restriction. For researchers, the technical report details on LSA and MOPD are worth reading as examples of how to engineer around the specific bottlenecks of very large, very long-context MoE models.