LongCat 2.0 and the Chinese Giants Shipping Frontier AI

Meituan LongCat 2.0 is the strangest sentence in AI this summer: a food-delivery super-app now ships a 1.6-trillion-parameter open-weight model that, by its maker's own testing, edges past GPT-5.5 on a serious software-engineering benchmark. Unveiled June 30, 2026 on GitHub, Hugging Face, and Meituan's own platform, it turned out to be the engine behind "Owl Alpha" — the anonymous stealth model that had been topping OpenRouter's developer charts.

The bigger story is the pattern, not the model. Meituan delivers food. Ant Group runs Alipay. Tencent runs WeChat and the world's largest games business. None of them is an AI-native lab — and all three now publish open-weight, frontier-adjacent models. That is a different phenomenon from DeepSeek, Qwen, Moonshot, MiniMax, or Z.ai shipping open weights, because those companies were built around model development from day one.

This post covers what LongCat 2.0 actually is, the honest read on its Meituan-reported benchmarks, where it sits on price, and — because the naming collisions are genuinely confusing — a clean disambiguation of who ships what: LongCat is Meituan, Ling is Ant Group, Hunyuan is Tencent.

For roughly two months before the reveal, developers routing traffic through OpenRouter kept picking an anonymous model called "Owl Alpha." By reveal time it accounted for roughly 10.1 trillion monthly tokens on the platform — a 242% month-over-month jump that put it in OpenRouter's global top three — and, per VentureBeat's reading of the platform data, it ranked first on the Hermes Agent workspace, second across Claude Code deployments, and third in international OpenClaw environments. Those are third-party usage observations rather than audited metrics, but the direction was unambiguous: an unknown model was winning real agentic workloads on merit.

On June 30, 2026, Meituan ended the mystery, publishing LongCat 2.0 on GitHub, Hugging Face, and its own platform, with OpenRouter listing the model under Meituan's own name in the weeks after. Three weeks on, the release reads less like launch-day news and more like a marker for a structural shift — which is how we treat it here.

LongCat 2.0 did not appear from nowhere. Meituan's first open-source LLM, LongCat-Flash, shipped September 1, 2025 — 560B total parameters, ~27B active, MIT license, 128K context — followed later that year by LongCat-Flash-Thinking, a dedicated reasoning variant. Roughly ten months separate Flash from LongCat 2.0, and the jump is stark on every axis:

LongCat 2.0 is a sparse Mixture-of-Experts model: 1.6 trillion total parameters, with roughly 48 billion active per token — and that activation count is dynamic, ranging from 33B to 56B depending on query complexity. A design Meituan calls Zero-Compute Experts routes simple tokens through near-free subnetworks, so the model spends its compute budget where the input actually demands it.

The million-token context comes from LongCat Sparse Attention (LSA), which Meituan describes as an evolution of DeepSeek Sparse Attention. LSA combines three indexing mechanisms — Streaming-aware Indexing, Cross-Layer Indexing, and Hierarchical Indexing — to keep long-context cost closer to linear than quadratic. A third, quieter piece is an N-gram Embedding module inherited from the smaller LongCat-Flash-Lite line: 135B parameters sitting orthogonal to the MoE expert layout, expanding embedding capacity roughly 100x to capture local token relationships and cut memory I/O during large-batch decoding.

Post-training is also unconventional. Meituan's MOPD pipeline — Multi-Teacher Optimization via Mixture of Specialized Experts — segregates fine-tuning into three isolated clusters (Agent Experts, Reasoning Experts, Interaction Experts) that a dynamic gate-routing mechanism fuses back together at inference time.

The claim that will age most interestingly has nothing to do with benchmarks. Per VentureBeat's reporting, LongCat 2.0 was trained and served entirely on domestic Chinese AI ASICs — a cluster of more than 50,000 accelerators, with no Nvidia GPUs in either the training or inference path. The pretraining run covered 35+ trillion tokens. None of the sources we reviewed names the chip vendor, so neither will we.

The contrast with the model's own lineage makes the point sharper than any policy essay could: LongCat-Flash, ten months earlier, was trained on Nvidia H800s. In one generation, Meituan moved a frontier-scale training run off American silicon entirely — and shipped the result under an MIT license for anyone to download.

Meituan positions LongCat 2.0 as an agentic coding model, and its headline claim is narrow and specific: on SWE-bench Pro, LongCat 2.0 scores 59.5 against GPT-5.5's 58.6 — per Meituan's own comparison. The company is more modest elsewhere, acknowledging that the model trails closed frontier systems like Claude Opus 4.8 on broad general-agent benchmarks such as FORTE and BrowseComp, while edging ahead specifically on software-engineering tasks.

The honest interpretation: a 0.9-point margin on one benchmark, self-reported, is not proof of frontier parity. What the OpenRouter usage data adds, though, is a second signal the benchmarks cannot fake — millions of real routing decisions by developers who did not know whose model they were using. When a stealth model climbs to a platform's top three on blind merit and then turns out to be open-weight and cheap, that says more about practical usefulness than any vendor table.

Standard pay-as-you-go pricing is $0.75 per million input tokens and $2.95 per million output tokens, per OpenRouter and Meituan's own API docs. A limited-time launch promotion cuts that to $0.30 / $1.20, with context-cache hits processed free of charge. VentureBeat's launch-day pricing table places the promo rate in the same tier as MiniMax M3, another 1M-context open-weight agentic model at an identical $0.30 / $1.20 — just above Xiaomi's MiMo-V2.5 Flash ($0.10 / $0.30) and DeepSeek-V4-Flash ($0.14 / $0.28), and far below closed-frontier list rates like GPT-5.6 Sol ($5 / $30) or Claude Fable 5 and Mythos 5 ($10 / $50).

Access surfaces are broad for a three-week-old public model: the LongCat API Platform exposes both OpenAI-compatible and Anthropic-compatible endpoints under the model ID LongCat-2.0; OpenRouter lists it; and third-party agent harnesses — Claude Code, OpenClaw, OpenCode, and Codex among them — already support it. Early-July coverage noted full weights were still pending, but the Hugging Face model card now hosts the Safetensors weights under an explicit MIT license, so self-hosting is live as of late July.

The single most common error in coverage of this space is conflating LongCat (Meituan) with Ling (Ant Group). They are entirely separate lineages from entirely separate companies. Ant Group — the Alipay fintech giant — publishes the Ling family, which includes Ling-1T and Ling-2.6, under its AGI organization. Meituan — food delivery — publishes LongCat. The names share nothing but a country of origin and an open-weight distribution strategy.

The third giant is Tencent, whose Hunyuan Hy3 open-weight reasoning model fully released on July 6, 2026 — 295B total parameters, 21B active, 256K context, Apache 2.0, listed on OpenRouter at roughly $0.20 / $0.80 per million tokens. We have also covered how Hy3's reasoning-effort routing works for agentic tasks in depth. Put the three side by side and the pattern this post is about becomes visible:

A note on the Ant Group row: the Ling family predates this post and spans multiple releases, and we have deliberately left cells blank rather than quote sizes and terms we could not verify against primary sources during research for this piece. The row exists for disambiguation — Ling belongs to Ant, not Meituan — not as a spec sheet.

Each giant has its own business logic. Meituan — founded in March 2010 by Wang Xing as a Groupon-style deals site, merged with Dianping in 2015, now a Hong Kong-listed super-app with 770M+ annual transacting users and 14.5M+ merchants — faces domestic margin compression and has publicly committed to investing "billions" into AI and domestic chip capability. Ant's fintech-scale agent workloads and Tencent's WeChat-ecosystem agentic tooling supply their own motives. Most coverage treats each launch in isolation; connected, they read as the same strategic bet — that owning frontier-adjacent model capability is now table stakes for operating a Chinese platform business, and that open-weight distribution is the fastest route to developer adoption.

The demand side explains why the bet is working. OpenRouter's early-July snapshot (week of July 8, 2026) shows Chinese open-weight labs dominating token volume — DeepSeek alone at roughly 5.4T tokens per week, with Xiaomi, MiniMax, Z.ai, and Tencent all in the weekly leaders, and Anthropic the only Western lab holding multiple top-10 slots. The steepest growth that week: Tencent +37% on the Hy3 launch, DeepSeek +26%, Z.ai +25%. And the spend mix shows where the money is going:

Agent and code workloads together account for well over half of spend — exactly the segments where cheap, high-context Chinese MoE models keep taking share. There is a geopolitical reading too, which we attribute rather than endorse: VentureBeat argues the release lands precisely as Washington pressures top-tier American labs to restrict access to their latest models — citing OpenAI being required to limit GPT-5.6 access and Anthropic taking Claude Fable 5 and Mythos 5 offline following U.S. government orders in the same period — which in its analysis has "left a wide operational window" for affordable Chinese open-weight alternatives. That is VentureBeat's editorial framing, not a Meituan statement, but the OpenRouter share data is consistent with it.

For teams building agentic systems, LongCat 2.0 changes the evaluation shortlist rather than the production default. The decision logic we would apply:

Looking forward, the wave logic matters more than any single model. Three non-AI-native giants have now demonstrated that a platform company with enough compute and data can ship a credible open-weight frontier-adjacent model — and the OpenRouter share numbers suggest developers will adopt them on price and merit within weeks. If that holds, the second half of 2026 should bring more entrants, faster price erosion in the open tier, and growing pressure on closed labs to justify a 10x-plus price premium on workloads where vendor benchmarks keep converging. We ran the same comparison logic on how MiniMax M3 stacks up against frontier closed models — the pattern repeats. For teams that want help building the eval harness rather than the hot take, our AI transformation engagements start with exactly this kind of model-routing decision.