Meituan released LongCat-2.5 on Monday — a 1.6-trillion-parameter multimodal AI model built for long-horizon autonomous coding tasks — and trained the entire system on domestically produced Chinese AI accelerators for the second time in a row. The release is today’s clearest evidence that US export controls on Nvidia chips have not prevented Chinese labs from reaching frontier-adjacent performance — but the architectural reason why is the story the headlines keep missing.
What Mixture-of-Experts Actually Does for China’s Compute Deficit
LongCat-2.5 is a mixture-of-experts model: a neural network architecture in which the total parameter count and the active-inference parameter count are two entirely different numbers. LongCat-2.5 carries 1.6 trillion total parameters but activates only approximately 48 billion of those for any given input token. That 48 billion figure is what determines inference cost.
This distinction is the architectural bridge that makes Chinese frontier AI viable on non-Nvidia hardware. Meituan’s training cluster consisted of roughly 50,000 domestically produced AI accelerators; the community has identified the likely chip as Huawei’s Ascend 910C, though Meituan’s own announcement uses the phrase “alternative hardware platforms” without naming a vendor. The Ascend 910C delivers approximately 60% of an H100’s per-chip performance; per chip, China’s domestic accelerator falls short. But MoE architecture routes each token through only a small expert sub-network rather than the full model, so the per-chip memory and compute gap matters far less than it would for a dense model of equivalent scale. The export controls worked as intended at the hardware layer — the Ascend 910C is genuinely inferior to an H100 per device — but MoE sparse activation dissolved the significance of that hardware disadvantage at the model layer.
Meituan’s key architectural innovation is LongCat Sparse Attention, or LSA. LSA addresses two specific failure modes in DeepSeek V3’s Lightning Indexer: output discontinuity and a quadratic scoring bottleneck. It operates through three components: Streaming-aware Indexing, which reshapes the token-selection budget for contiguous memory reads; Cross-Layer Indexing, which reuses one indexing pass across two adjacent layers; and Hierarchical Indexing, which performs coarse block-level recall before fine token selection. The model also carries 135 billion N-gram Embedding parameters — a sparse block orthogonal to the MoE routing that Meituan says prevents performance degradation where standard sparse activation would otherwise reach its limits.
Pretraining ran across over 35 trillion tokens and reportedly completed without a single rollback or irrecoverable loss spike — a meaningful engineering achievement for a run of this scale on hardware with significantly less per-device memory than an H800.
What LongCat-2.5 Is Actually For
The model targets long-horizon autonomous agent tasks: complex coding jobs requiring an AI to maintain coherent, goal-directed behavior across dozens to hundreds of dependent sequential steps. This is distinct from a question-answering chatbot. LongCat-2.5 is designed to plan, write code, execute it in a terminal or browser, analyze the result, recover from failures, and deliver a finished artifact — without a human managing each individual step.
To support sustained autonomous operation, LongCat-2.5 retains the 1-million-token context window first introduced in LongCat-2.0. One important caveat from independent analysis of LongCat-2.0, which shares the same architectural lineage: the shipped model configuration caps context at 256K, with extension to 1 million provisioned but not active by default. The 1M figure describes the training data distribution, not the immediate serving default.
LongCat-2.5 adds native multimodality — image understanding built into the token framework — and extends the model’s commercial analysis and software development capabilities for Meituan’s local-services business. Meituan deployed the model inside Miaoshow (妙手), an AI merchant management workbench that helps restaurant and hotel operators on its platform analyze business performance, understand industry benchmarks, and execute operational tasks within their authorized scope.
Meituan also incorporated a failure-recycling loop in training: agent breakdowns from production deployments were converted into training tasks, allowing the model to learn from real-world failure modes rather than only from curated evaluation sets.
What the Benchmarks Say — and What They Don’t
Meituan’s announcement states that LongCat-2.5 performs between Claude Opus 4.8 and Claude Opus 5 across several mainstream evaluations, surpassing Claude Opus 5 on select tasks. That claim carries a significant caveat: it is entirely self-reported, from in-house testing, and no independent evaluator has published a third-party assessment of LongCat-2.5 as of this writing — a pattern consistent with the predecessor model.
For context on what “self-reported” has meant for this model family: BenchLM, a third-party comparison aggregator, found zero shared benchmarks between GPT-5.5 and LongCat-2.0 as of August 2026, concluding that “the public evidence does not support any quality verdict” on head-to-head comparisons. Separately, Artificial Analysis’s independent index placed LongCat-2.0 at 45.3 on its Coding Index — better than roughly 49% of models — which is a notably different impression than the headline SWE-bench Pro score of 59.5 on the official model card. The SWE-bench Pro edge Meituan claimed over GPT-5.5 (59.5 versus 58.6) was itself conducted in-house.
Independent developers who ran billions of tokens through LongCat-2.0 during its two months operating anonymously as “Owl Alpha” on OpenRouter characterized it as a “dependable workhorse” in community field reports: genuinely strong at instruction following, planning, and maintaining coherence at high contexts, but not matching frontier models on raw reasoning benchmarks. The official benchmark table confirms that shape: LongCat-2.0 led on coding and agent benchmarks specifically and trailed Claude Opus 4.8 by approximately 9.7 points on SWE-bench Pro.
For LongCat-2.5, the full picture will only become clear once external evaluators receive access. Until then, “between Opus 4.8 and Opus 5” should be treated as a directional claim, not a settled result.
Who Is Running This — and the Data Governance Gap
Developers can access LongCat-2.5 through Meituan’s LongCat API platform at longcat.ai, or through OpenRouter international access path, which has served the bulk of the model family’s international traffic. LongCat-2.0’s launch API pricing was $0.30 per million input tokens and $1.20 per million output tokens (a discounted rate; list price was $0.75/$2.95), with no context-length surcharge — a meaningful cost advantage for repository-scale work where competing models typically increase prices above 200K tokens. Pricing for LongCat-2.5 was not yet independently confirmed at time of writing.
But API pricing is only one dimension of the evaluation decision. For enterprise buyers, it is typically not the first.
Meituan’s API platform FAQ does not state a data retention policy. It does not disclose whether prompts or completions are used for model training. It does not name the physical country of its inference servers. No service-level agreement, enterprise data processing agreement, or public security audit of LongCat’s infrastructure was available as of this writing. An independent review from eesel AI in August 2026 identified this as the primary production-deployment blocker: teams handling personal data, regulated industry data, or sensitive code face questions the platform documentation simply doesn’t answer.
What China’s Law Requires — Regardless of the Privacy Policy
Independent of Meituan’s stated policies, a structural legal condition applies to any Chinese company’s data handling that cannot be contracted around: China’s National Intelligence Law (2017), Article 7, requires that “all organizations and citizens shall support, assist, and cooperate with national intelligence work in accordance with law.” This statutory obligation applies to Meituan as a Beijing-headquartered company regardless of where its API servers are physically located, what the privacy policy states, or what its terms of service promise.
Related provisions in China’s Data Security Law (2021) and Cybersecurity Law (2016) add data localization requirements and government-access provisions. Legal scholars debate the precise operational scope of these laws — Beijing Foreign Studies University’s Gu Bin has argued Article 7 does not authorize pre-emptive data collection; defense analyst Murray Scot Tanner writing in Lawfare argued it converted citizen obligations from intelligence “defense” to “offense” — but no published interpretation concludes that Chinese companies are exempt from these obligations when operating commercial APIs. Enterprise buyers should treat this as a fixed legal condition of deploying Chinese-origin API services, not a risk factor to weigh against pricing. The ASPI Strategist has documented the broader ambiguity of China’s intelligence and counter-espionage laws in the context of commercial technology deployments.
Performance Gaps and Where the Model Still Trails
Based on LongCat-2.0’s public data, which represents the only benchmarked baseline until LongCat-2.5 independent assessments arrive, the model trails Claude Opus 4.8 on every benchmark where comparison data is available. The gaps are largest on GPQA-diamond (92.4 for Opus 4.8 versus 88.9 for LongCat-2.0), on retrieval benchmarks (BrowseComp: 84.3 versus 79.9; FORTE: 77.2 versus 73.2), and on web research (RWSearch: 77.3 versus 78.8, where LongCat-2.0 actually edges ahead). The honest summary: this is a strong coding-and-agent specialist, not a model that leads the current generation across the board.
Self-hosting is a significant practical constraint. LongCat-2.0’s weights, the baseline template for LongCat-2.5, require a minimum of 8 Nvidia B300 GPUs or 16 H200-class GPUs across two nodes for the only verified serving recipe. No GGUF format exists, meaning Ollama, LM Studio, and consumer-grade local AI setups cannot load the model. The BF16 weight file runs 3.55 terabytes; the FP8 quantized version 2.05 terabytes. Open-weight releases for LongCat-2.5 on GitHub and Hugging Face are planned but had not yet been published at the time of this article’s publication.
Ecosystem and Operational Friction
For developers integrating the LongCat API, the documented parameter surface is narrow: seven parameters supported, no JSON mode, no stop sequences, no seed. Tool calling appears in the parameter list but the schema is undocumented; at LongCat-2.0’s launch, developers reported that the model used a proprietary
Payment for LongCat API credits is documented only through WeChat Pay and Alipay — mainland China accounts — which means most developers outside China must route through OpenRouter or similar services.
The Huawei Chip Complication
The Huawei Ascend 910C chips that Meituan’s training cluster almost certainly relied on carry a separate US export control issue. In May 2025, the US Commerce Department’s Bureau of Industry and Security issued guidance on Huawei Ascend chips stating that their use “anywhere in the world” violates US export controls, given that those chips were “likely developed or produced in violation of US export controls.” This does not prevent international developers from using LongCat-2.5’s API, but it is relevant context for evaluating the broader claim that this training run operates entirely outside the scope of US regulatory interest.
Who LongCat-2.5 Is Genuinely For
LongCat-2.5 is the most capable model yet from a company that built its reputation delivering dumplings and became a credible frontier AI lab in roughly three years. For developers running high-volume, long-context, agentic coding work on non-sensitive repositories, it represents a serious option — assuming its self-reported benchmark performance holds up under independent testing, and assuming the data governance documentation gaps are acceptable to the team’s security posture.
The mandatory evaluation framework is four-dimensional: the performance gap that still separates it from current-generation models like Claude Opus 4.8 on general reasoning and retrieval; the absence of any independent third-party benchmarks for this specific release; the ecosystem friction that creates practical obstacles for teams outside China and those requiring standard enterprise contractual protections; and China’s National Intelligence Law, a fixed legal condition that applies regardless of pricing, privacy policy, or server geography. For developers for whom those four conditions are acceptable, the architecture is real, the production history is real, and the model’s particular strength in long-horizon coding coherence makes it genuinely worth evaluating. For anyone for whom one of those four dimensions is a dealbreaker — and enterprise security teams will almost certainly flag the fourth — that remains true regardless of what the benchmarks eventually show.
Frequently Asked Questions
How does a mixture-of-experts architecture make frontier-scale AI training possible on non-Nvidia hardware?
In a mixture-of-experts design, total parameter count and active inference compute are two separate numbers. LongCat-2.5 has 1.6 trillion total parameters but activates only approximately 48 billion per input token — that smaller number determines inference cost. China’s Huawei Ascend 910C chip delivers roughly 60% of an Nvidia H100’s per-chip performance, which is a real hardware gap. But MoE’s sparse activation dissolves much of that gap’s practical significance: you’re not asking each chip to process the full model, only to serve a small expert sub-network for each token. This is the architectural bridge that made training at scale on domestic Chinese silicon meaningful, despite the per-chip performance disadvantage.
Have any third parties independently tested LongCat-2.5’s benchmarks?
No. As of November 2, 2026, all performance claims for LongCat-2.5 come from Meituan’s own in-house testing. The pattern is consistent with the predecessor: as of August 2026, BenchLM found zero shared third-party benchmark results between GPT-5.5 and LongCat-2.0, explicitly stating that “the public evidence does not support any quality verdict.” Meituan’s claim that LongCat-2.5 performs between Claude Opus 4.8 and Opus 5 should be read as a directional claim pending independent evaluation.
What does China’s National Intelligence Law mean for developers considering the LongCat API?
China’s National Intelligence Law (2017), Article 7, requires all Chinese organizations to support, assist, and cooperate with state intelligence work on demand. This applies to Meituan as a Beijing-headquartered company regardless of where its servers are located, what its privacy policy says, or what its terms of service promise — it is a statutory condition of operating as a Chinese company, not a contractual variable. LongCat’s API platform FAQ also does not disclose data retention periods, whether prompts are used for model training, or the physical location of inference servers. Enterprise teams subject to SOC 2, GDPR, HIPAA, or standard corporate data handling reviews should treat these as threshold questions rather than incidental factors.
What are the practical limits of LongCat-2.5 for developer adoption outside China?
Based on LongCat-2.0’s documented architecture: self-hosting requires at minimum 8 Nvidia B300 GPUs or 16 H200-class GPUs — no consumer-friendly format like GGUF exists. The API parameter surface is narrow (seven parameters, no JSON mode, underdocumented tool calling). Payment for API credits on Meituan’s own platform is documented only through WeChat Pay and Alipay, so most developers outside China rely on OpenRouter as the access path. Open-weight releases for LongCat-2.5 specifically are planned but were not yet published at the time of this article.
