This deep-dive review, published the day after launch, establishes the most important background fact — the anonymous free model "Owl Alpha," which ran on OpenRouter for two months, was LongCat-2.0 (processing approximately 10 trillion tokens per month and ranking first on the Hermes Agent leaderboard) — while breaking down the official claims one by one: the narrow SWE-bench Pro win over GPT-5.5 was an official self-test, the weights were still only "planned" when the review was published, and the open-source field's true rival was GLM-5.2 (62.1 vs. 59.5).
Owl Alpha ran anonymously on OpenRouter for two months; when developers called it, they did not know the vendor or internal structure. At launch it was processing approximately 10 trillion tokens per month, more than any model on OpenRouter's Hermes Agent leaderboard.
The only evidence not self-reported by the official source was the two months of anonymous usage: "This is the strongest piece of evidence in the entire launch, and the only thing Meituan did not generate itself."
Risk disclosure: Owl Alpha's OpenRouter listing disclosed that prompts and completions might be logged by the provider to improve the model — free users unknowingly supplied training signals for two months.
Zero-Computation Experts: simple tokens (variable names, parentheses) take an almost empty computation path, while difficult tokens go through the full expert stack; routing is per token rather than per request.
LSA: the official source acknowledged that the indexer itself became a new bottleneck (scoring every layer and token remains quadratic, and the memory-access pattern does not match hardware prefetching); SI/CLI/HI are the three corresponding fixes. The risk is that "when incorrect routing sends a token that needs full reasoning down a cheap path, it will fail silently," and no public figure explains how often this happens.
N-gram Embedding: n-gram size 5, approximately 128B parameters (official figure: 135B), and an effective vocabulary expansion of approximately 900x; MoE + N-gram produces effective sparsity of approximately 97%, with N-gram accounting for <10% (returns diminish after >30%).
MOPD: post-training is split into three expert groups — Agent / Reasoning / Interaction — with a gating network routing by task at inference time — "probably the most genuinely novel part"; most labs only merge weights by averaging.
The official source says this is the first trillion-parameter model trained entirely on a domestic cluster (approximately 50,000 accelerators, no Nvidia), but does not name the chip vendor; the official acknowledgment of Huawei's HCCL communication library and independent estimates point to Huawei Ascend 910C.
Official engineering claims: monthly hardware failure rate reduced by 70%+ and 35T tokens processed with no rollback or unrecoverable loss spikes — "achieving this on immature non-CUDA hardware is a real accomplishment."
In the official self-test table, the only model LongCat-2.0 consistently beats is Gemini 3.1 Pro; against GPT-5.5 it only narrowly wins SWE-bench Pro (59.5 vs. 58.6) and trails across the rest (FORTE 73.2 vs. 77.8, RWSearch 78.8 vs. 85.3, BrowseComp 79.9 vs. 84.4).
Claude Opus 4.8 is a bigger threat to LongCat-2.0: on the same SWE-bench Pro test it scores 69.2 (nearly 10 points ahead), while Opus 4.7 also scores 64.3.
Three claims thin out under line-by-line review:
The weights had not been released at the time — GitHub explicitly said "Model weights coming soon, stay tuned" (only 76 stars/4 forks 10 hours after launch); media reports that "Meituan open-sourced" it described a plan rather than a release. (Note: as of 2026-08-18, the weights have been released on HF, so this point is outdated.)
The strongest open-source rival, GLM-5.2, was absent from the official comparison chart: GLM-5.2 scored 62.1 > 59.5 on the same SWE-bench Pro test, and its MIT-licensed weights were downloadable two weeks earlier.
The official chart used the company's own harness and scoring throughout, and noted that "problematic tasks [were] corrected"; at publication, there were no independent third-party scores from Artificial Analysis, Scale, or anywhere else.
Cache hits are most cost-effective for "Agents that repeatedly reread the same context" (single-repository coding Agents, single-long-document research Agents); one-shot prompts get almost no benefit.
DeepSeek V4-Pro's list price is still lower than LongCat-2.0's; use GLM-5.2 to verify the benchmarks (downloadable weights, higher SWE-bench Pro score); browsing, multi-step tool chains, and in-task judgment remain areas where closed frontier models lead.
The article was published on 2026-07-01 (the day after launch). "Weights not released" and some pricing data ($0.69/$2.78 per 1M) are outdated: the weights are now available on HF (Prompt Directory 02), official current pricing is ¥5/¥20 (Prompt Directory 01), and OpenRouter's current price is $0.30/$1.20 (Review 03).
The article is analytical commentary rather than a controlled evaluation. Its figures are relayed from official and platform sources, but key data such as "10 trillion tokens/month" and "GLM-5.2 62.1" have no independent original-source links (the author says source links are in the first reply); treat them as second-hand when citing.
It is best read alongside the official documents (Reviews 01 and 02): the official documents provide the full details, while this article provides a critical framework.
LongCat 2.0
