Meituan's 1.6 Trillion-Parameter Model Now Runs on Chinese Chips
Meituan launched LongCat-2.5 on November 1 to 2, with 1.6 trillion total parameters and about 48 billion activated per token, long-context training and inference optimized for Chinese AI accelerators, company-reported benchmark scores, and weights not yet released.
- 01THE LAUNCH — On November 1, the LongCat team published the LongCat-2.5 technical announcement.
- 02THE RELEASE — On November 2, the model went fully live on the chat service and API.
- 03THE OPTIMIZATION — Long-context training and 4-bit inference optimized for Chinese accelerators.
WHAT HAPPENEDWhat Happened
Meituan's LongCat team launched LongCat-2.5 between its November 1 technical announcement and the November 2 API launch. It is a natively multimodal model built for long-horizon agent tasks, with 1.6 trillion total parameters and about 48 billion activated per token.
The selling point is not the parameter count but the training base. Meituan says long-context training and 4-bit inference are optimized for Chinese AI accelerators, with memory per Chinese accelerator under 80GB, so the training system prioritizes memory efficiency.
The model is available through Meituan's chat service and API, with a 1 million-token context window, up to 128,000 tokens of output, and a limited-time price of $0.30 per million input tokens. GitHub and Hugging Face releases are marked as coming soon and are not yet live.
BIG TOTAL, SMALL ACTIVEBig Total, Small Active: Read the Two Numbers Separately
1.6 trillion is the total parameter count; 48 billion is what is actually activated per token, only 3% of the total. That is the point of a mixture-of-experts architecture: the model can be very large without each call costing proportionally more.
Meituan's 4-bit quantization scheme compresses both expert weights and activations to 4 bits. The company says this configuration tested on par with full-precision models. The memory and compute saved are what make running a large model on Chinese chips workable.
SELF REPORTED SCORESSelf-Reported Scores Must Be Read as Self-Reported
In Meituan's own comparison table, LongCat-2.5 scored higher on 20 of 27 tests shared with GPT-5.6-Sol, and on 10 tests shared with Claude Opus 5. These numbers come from Meituan's own testing, with some rival scores taken from public reports or leaderboards.
Independent verification does not exist yet. The weights have not been released, so third parties cannot reproduce the results. Reading self-reported scores as an industry ranking is the easiest mistake in this story. The scorecard's value is directional, not ordinal.
THE DOMESTIC CHIP LINEThe Domestic-Chip Line Keeps Getting More Solid
LongCat-2.0 already walked this path in June with Chinese-chip training, officially using more than 50,000 Chinese accelerators and no NVIDIA hardware, with weights later open-sourced under the MIT license. LongCat-2.5 is the second step on the same line.
The significance is not a single breakthrough but repeatability. The same optimization approach carried from one generation to the next shows that running large models on Chinese accelerators has moved from experiment to process. Once a process holds, scale is a matter of time.
THE SHIFTWhat Changed
SIGNAL JUDGMENT — ANALYSISRunning big models on Chinese chips has gone from experiment to process
Tokenta Signal's judgment: the real signal in LongCat-2.5 is not the parameter count but the repeatability of a second generation of Chinese-chip training. Once a process runs twice, scale is a matter of time.
WHAT REMAINS UNCERTAINWhat Remains Uncertain
EVIDENCEEvidence
The original confirms Meituan launched LongCat-2.5 on November 1 to 2, with 1.6 trillion total parameters and about 48 billion activated per token, long-context training and 4-bit inference optimized for Chinese AI accelerators, and company-reported benchmark scores.
PRIMARY SOURCESPrimary Sources
- FourWeekMBA: Meituan's 1.6T Model on Chinese Chips Rivals OpenAI GPT-5.6
https://fourweekmba.com/ai-meituans-1-6t-model-on-chinese-chips-rivals-openai-gpt-5-6/ - Tech Times: Meituan Releases LongCat-2.5 on Chinese Chips; Benchmark Claims Await Independent Tests
https://www.techtimes.com/articles/329925/20261103/meituan-releases-longcat-25-chinese-chips-benchmark-claims-await-independent-tests.htm
RELATEDRelated Intelligence
- Microsoft Winds Down Its Oracle AI Cloud Rental as Older Nvidia Chips Fall Behind
- An Unreleased Anthropic Model Proved an 18-Year-Old Quantum Conjecture, Machine-Checked in Lean
- IBM Unveils Quantum System Two in Chicago, Its Most Advanced Computer Yet, Powered by Nighthawk
- Vistra Posts $1.14B Q3 Profit as Power EBITDA Jumps Over 50%, Lifts 2026 Outlook