Back to list
Meituan LongCat Open-Sources General 365: A Rigorous New Benchmark for AI Reasoning Performance
Industry NewsMeituanAI BenchmarkingReasoning Models

Meituan LongCat Open-Sources General 365: A Rigorous New Benchmark for AI Reasoning Performance

Meituan's LongCat team has officially released General 365, a new open-source benchmark designed to evaluate the reasoning capabilities of large language models (LLMs). The benchmark's debut has sent ripples through the AI community by revealing a significant performance gap in current technology. In a comprehensive test of 26 mainstream models, even the industry-leading Gemini 3 Pro managed an accuracy rate of only 62.8%. More strikingly, the vast majority of the models tested failed to reach the 60% threshold, which is typically considered a passing grade. This release by Meituan Technical Team establishes a new, more challenging standard for AI reasoning, suggesting that current models still face substantial hurdles in complex cognitive tasks.

美团技术团队

Key Takeaways

  • New Evaluation Standard: Meituan's LongCat team has launched General 365, an open-source benchmark specifically focused on AI reasoning.
  • Gemini 3 Pro Performance: The model currently regarded as the strongest, Gemini 3 Pro, achieved an accuracy of 62.8% on the benchmark.
  • Widespread Failure to Pass: Most of the 26 mainstream models tested failed to reach a 60% accuracy score, highlighting a significant deficiency in current reasoning capabilities.
  • Industry Benchmark Shift: General 365 aims to set a more rigorous bar for evaluating how models handle complex reasoning compared to existing metrics.

In-Depth Analysis

The Launch of General 365 and the Reasoning Crisis

The release of General 365 by Meituan's LongCat team marks a pivotal moment in the evolution of AI benchmarking. For years, the industry has relied on a variety of metrics to measure the progress of large language models. However, as models become more sophisticated, many existing benchmarks have begun to suffer from saturation, where top-tier models achieve near-perfect scores, making it difficult to distinguish true reasoning ability from pattern matching or data memorization.

General 365 addresses this by introducing a framework that appears significantly more demanding than its predecessors. By open-sourcing this tool, Meituan is providing the global developer community with a "reality check." The initial data provided by the LongCat team suggests that the industry is currently facing a "reasoning crisis." When 26 of the most prominent models are put to the test and the majority cannot even secure a 60% accuracy rate, it indicates that the path toward true artificial general intelligence (AGI) is still fraught with fundamental challenges in logical processing and multi-step reasoning.

Analyzing the Performance of Gemini 3 Pro

The most telling data point from the General 365 release is the performance of Gemini 3 Pro. As a model widely recognized as one of the most capable in the world, its score of 62.8% serves as a benchmark for the current "ceiling" of AI reasoning. While 62.8% represents the top of the class in this specific evaluation, it is a modest figure in absolute terms.

This score suggests that even the most advanced architectures are struggling with the specific types of reasoning tasks curated in General 365. The fact that the "strongest" model is only slightly above the 60% mark implies that General 365 is designed to expose the edge cases and complex logical dependencies where current LLMs typically fail. It shifts the narrative from how well models can generate text to how accurately they can navigate complex problem-solving environments. For researchers, the 62.8% mark is not just a score; it is a target that defines the current frontier of the industry.

The 60% Threshold: A New Baseline for AI Maturity

Perhaps the most alarming revelation from the LongCat team's report is that the vast majority of mainstream models failed to reach the 60% accuracy threshold. In many academic and professional contexts, 60% is the baseline for a passing grade. The failure of most models to reach this level on General 365 suggests that many current AI solutions may be less reliable in high-stakes reasoning scenarios than previously thought.

This widespread underperformance highlights a potential over-optimization of models for conversational fluency at the expense of deep reasoning. As Meituan sets this new "ruler" for the industry, it forces a re-evaluation of what constitutes a "capable" model. If a model can write poetry but cannot pass a basic reasoning threshold on General 365, its utility in technical, legal, or scientific fields may be limited. The benchmark effectively separates models that are merely good at language from those that possess genuine analytical depth.

Industry Impact

The introduction of General 365 is likely to influence the AI industry in several key ways. First, it provides a transparent, open-source metric that discourages "benchmark gaming," as the difficulty level is high enough to reveal true performance variances. Second, it places pressure on major AI labs to improve the logical consistency of their models rather than just increasing parameter counts or training data volume.

Furthermore, Meituan’s decision to open-source the benchmark allows smaller research teams to align their development with industry-leading standards. By identifying that even the best models are currently hovering around the 60% mark, General 365 defines the next phase of the AI arms race: the quest for robust, reliable reasoning. This will likely lead to a shift in training methodologies, with a greater emphasis on synthetic reasoning data and reinforcement learning from human feedback (RLHF) focused on logical accuracy.

Frequently Asked Questions

Question: What is General 365?

General 365 is an open-source reasoning evaluation benchmark released by the Meituan LongCat team. It is designed to test the complex reasoning capabilities of mainstream large language models through a rigorous set of evaluations.

Question: How did the top AI models perform on this benchmark?

According to the report, Gemini 3 Pro, currently considered the strongest model, achieved an accuracy of 62.8%. However, the majority of the 26 mainstream models tested failed to reach an accuracy of 60%.

Question: Why is the 60% score significant in this context?

The 60% score is significant because it is often viewed as the minimum threshold for a "passing" grade. The fact that most models failed to reach this mark suggests that current AI technology still has a long way to go in mastering complex reasoning tasks.

Related News

How Google Duplex Evolved to Call Businesses on Your Behalf in Gemini
Industry News

How Google Duplex Evolved to Call Businesses on Your Behalf in Gemini

A Google Research Blog entry titled “How Google Duplex evolved to also call businesses for you in Gemini” describes its subject as the evolution of Google Duplex to support calls to businesses on a user’s behalf in Gemini. The entry is dated November 20, 2026, and the supplied article text consists only of the label “Generative AI.” The headline identifies the capability and connects it to Duplex, but the available text does not explain how the system evolved, how users request a call, or where the feature is available. The headline establishes the topic, but the supplied material does not provide a technical account or rollout details.

How AI and Lean Are Helping Automate Formal Verification to Fight Software Slop
Industry News

How AI and Lean Are Helping Automate Formal Verification to Fight Software Slop

In Latent Space, Richard MacManus examines formal verification as an emerging way to fight what the headline calls “slop.” The central claim is that proving software correct with mathematics has traditionally required specialist work, and that AI and Lean are now helping automate that process. The distinction is important: the reported change concerns how the work of proving correctness is done, not what counts as a proof. The supplied excerpt presents automation as assistance, without saying that it replaces specialists or completes proofs independently. It also does not specify which parts of the process AI and Lean help automate, how much work they save, or which software has been verified this way.

AI and Human Claim Solutions to the 167-Year-Old Riemann Hypothesis
Industry News

AI and Human Claim Solutions to the 167-Year-Old Riemann Hypothesis

TechCrunch AI reports that AI claimed a solution to the Riemann hypothesis, a mathematical problem the headline describes as 167 years old. A human also put forward a competing solution. The report, written by Tim Fernholz and published on November 20, 2026, describes mathematicians as astonished by the AI-driven result. The supplied excerpt does not explain either proposed solution, identify the AI system or the human claimant, or say whether either claim has been verified. The distinction between a claimed solution and a confirmed one is central to reading the report: the excerpt establishes that two claims are being reported, but it does not establish that either has resolved the hypothesis.