Meituan has released LongCat-2.5, a model with 1.6 trillion total parameters that the Chinese delivery company says it optimised for training and inference on domestic AI accelerators, according to the LongCat team’s launch post dated 1 November. In Meituan’s own benchmark table, LongCat-2.5 scored higher than OpenAI’s GPT-5.6-Sol on 20 of 27 tests where both models have a score, by our count.
The same table puts it above Anthropic’s Claude Opus 5 on 10 of 27 shared tests, by our count. The scores are self-reported: Meituan says rival scores marked with an asterisk come from public reports or leaderboards, and that the rest were run by Meituan and “are indicative only.”
The key insight: Meituan says its 1.6-trillion-parameter LongCat-2.5 was optimised for training and inference on domestic AI accelerators, and its own benchmark table puts it above OpenAI’s GPT-5.6-Sol on 20 of 27 shared tests and above Claude Opus 5 on 10, by our count.
LongCat-2.5 is “a native multimodal model for long-horizon agentic tasks,” the LongCat team wrote. It has about 48 billion parameters activated per token, according to the post. By our arithmetic, that is 3% of the 1.6 trillion total.
The team says it used “large-scale long-horizon reinforcement learning” to improve work on complex codebases and full-project builds. It also says it converts real-world agent failure trajectories into “reproducible, verifiable training tasks.”
The model is live in Meituan’s chat service and API, TechNode reported, with GitHub and Hugging Face releases planned to follow. The launch post marks both as “Coming soon.” Meituan’s API documentation lists a 1M-token context window with a maximum output of 128K tokens, and offers OpenAI- and Anthropic-format endpoints.
Benchmark scores Meituan reported for LongCat-2.0 and LongCat-2.5 on six agent benchmarks, from the LongCat-2.5 launch post of 1 November 2026. The scores are company-reported and have not been independently verified.
The Domestic Chips
“We optimize long-context training and W4A4 inference on domestic AI accelerators,” the LongCat team wrote. The post says each domestic accelerator used for long-sequence training has less than 80 GB of memory, so the training system “prioritizes memory efficiency.”
The post does not name the chip vendor. It says Meituan aligned its parallel groups with a “supernode” topology, keeping heavy traffic inside the high-bandwidth interconnect, and that its FSDP setup achieved 86% higher per-device throughput than a DDP baseline in long-sequence fine-tuning.
For inference, Meituan says it quantises both expert weights and activations to 4 bits. In its online stress tests, prefill throughput rose 43.3% and decode throughput 42.77% against an 8-bit baseline, the post says, with time to first token down 15.91%.
Business Pill · Sovereign AI
An evergreen Business Engineer explainer on sovereign AI. It teaches the general idea only and does not discuss Meituan or LongCat-2.5.
How It Compares
Meituan compared LongCat-2.5 with GPT-5.5, GPT-5.6-Sol, Claude Opus 4.8, Claude Opus 5, GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash across 31 benchmarks. Against its predecessor the jumps are large on agent coding tests: on Terminal Bench v4.0 the score went from 3.0 for LongCat-2.0 to 38.5, and on DeepSWE v1.1 from 9.7 to 64.6, according to the table.
Against the frontier models the picture is mixed. On Terminal Bench v2.1, LongCat-2.5 scored 88.4, against 88.8 for GPT-5.6-Sol and 89.1 for Claude Opus 5, the table shows. On BrowseComp it scored 91.9, against 90.4 and 90.8. On Terminal Bench v4.0, Claude Opus 5 scored 53.9.
By our count of the same table, LongCat-2.5 scored higher than DeepSeek V4.1 Flash on 12 of 26 shared tests. Meituan says it ran its coding tests for Terminal Bench and several others inside Anthropic’s Claude Code harness.
As we read the launch post, the claim has two parts: a frontier-scale agent model, and a training and inference stack tuned for China’s own chips, which the post describes but does not name.
As we read the table, the gap to the top models is narrow on some tests and wide on others: Meituan reports 38.5 on Terminal Bench v4.0 against 53.9 for Claude Opus 5.
The LongCat team, Meituan, in the LongCat-2.5 launch post
“We optimize long-context training and W4A4 inference on domestic AI accelerators”
Three Implications
FOR AI CHIP BUYERS IN CHINA Meituan says each domestic accelerator it used for long-sequence training has less than 80 GB of memory, and describes the parallelism and memory work it did around that limit.
FOR DEVELOPERS Meituan’s API documentation lists a 1M-token context window and OpenAI- and Anthropic-format endpoints, and the launch post says its coding tests ran inside Claude Code.
FOR MODEL RIVALS The launch post lists GitHub and Hugging Face releases as coming soon; until then, the benchmark results are Meituan’s own.
The framework argues that the physical layer matters as much as the model: “It’s a fundamental constraint on the speed of AI transformation, and it’s creating geopolitical leverage that operates independently of who has the best models.”
As we read it, Meituan’s launch post speaks to that layer: the company says it tuned training and inference for domestic accelerators. The framework is the owner’s argument and does not discuss Meituan.
The launch post does not name the domestic accelerators, and it describes long-context training and inference on them; it does not say every stage of training ran on them. The benchmark results are Meituan’s own and have not been independently reproduced. We could not load Meituan’s API pricing page, so this piece gives no price. The open-weight releases are listed as coming soon, with no date.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Meituan says it built a 1.6-trillion-parameter agent model with training and inference tuned for China’s own AI chips, and its own table puts it ahead of GPT-5.6-Sol on 20 of 27 shared tests and of Claude Opus 5 on 10. Until the weights and outside tests arrive, those are the company’s numbers.
A note on sourcing. Model details, infrastructure figures and benchmark scores are from Meituan’s LongCat-2.5 launch post of 1 November; availability is from TechNode and Meituan’s API documentation. The benchmark scores are company-reported. The 20-of-27, 10-of-27 and 12-of-26 counts and the 3% share are our arithmetic from Meituan’s table.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
