§ News
By AI Blog Editor
Nov 16, 2026 · 12 min read
A Chinese food-delivery company shipped a 1.6-trillion-parameter model trained on domestic chips — four days after Jensen Huang got fired
On Nov 2, 2026 Meituan launched LongCat-2.5, a 1.6T multimodal MoE. A week later it opened the weights under MIT. Meituan says it was trained on "domestic AI accelerators"; the self-hosting reference is 194 Huawei Ascend NPUs.

On Monday November 2, 2026, Meituan — the Chinese company most Westerners have never heard of and most Chinese commuters have used at lunch — launched LongCat-2.5, a native-multimodal mixture-of-experts model with 1.6 trillion total parameters and ~48 billion active per token. One week after that, Meituan opened the weights on Hugging Face under an MIT license. Crypto Briefing's launch write-up buries the structural fact in one sentence: "LongCat-2.5 was trained on domestically produced Chinese AI accelerators." The Decoder's Sunday follow-up puts a number on the self-hosting reference: 194 Huawei Ascend Atlas A2 chips, no Nvidia anywhere in the loop.
Jensen Huang was fired on Thursday November 12. LongCat-2.5's weights went public the Monday before. The timing was not a response — Meituan had the model queued for weeks — but one does not announce "AI chips are investable infrastructure" into a week in which a Chinese delivery app releases a trillion-parameter open-weights agent model trained without any of yours.
What Meituan actually shipped
LongCat-2.5 is a mixture-of-experts transformer. 1.6 trillion total parameters, ~48 billion active per token, 1-million-token input context, 128K output. Text in, images in, text out. Reasoning mode is a switch. The Hugging Face model card calls it "a native multimodal MoE model for long-horizon agentic tasks" — Meituan's product thesis in one noun phrase.
A W4A4 quantized version — four-bit weights, four-bit activations — ships alongside, and this is where the hardware-policy fingerprint shows. Meituan's own numbers put the W4A4 release at 43.3% faster prefill and 42.77% faster decode than the W8A8 MoE baseline on their deployment stack, which the model card ties to "native INT4 computation on domestic AI accelerators." That is a sentence written by a team whose chips run INT4 well and whose public message is that the quantization and the silicon were co-designed. If you are an export-control analyst at the US Commerce Department, that sentence is the piece of the launch you underline.
The benchmarks are self-reported, as benchmarks always are on release day. On Meituan's own tables LongCat-2.5 clears Terminal Bench v2.1 at 88.4, trailing Claude Opus 5 (89.1) and DeepSeek-V4.1-Flash (90.6); on BrowseComp it tops the table at 91.9; on the harder Terminal Bench v4.0 it drops to 38.5 against Opus 5's 53.9. As the model card politely puts it: "unmarked scores are measured in-house and are indicative only." Reading that twice is the correct response.
The API prices land, per The Decoder's testing, at ~€0.27 per million input tokens and €1.07 per million output tokens — introductory, promotional, time-limited. Claude Opus 5.5 charges roughly $75 per million output. This is not an apples-to-apples comparison, and everyone in the room knows that. Meituan is not trying to outsell Opus. It is trying to make the per-million-token line item for a Terminal-Bench-grade agent model look less like a quarterly budget decision and more like a rounding error.

The chips Meituan will not name
Meituan's technical material uses the phrase "domestic AI accelerators" four separate times and never once says "Huawei." The predecessor LongCat-2.0 (June 2026, same 1.6T/48B configuration without vision) was trained on more than 50,000 chips without Meituan ever naming the vendor, either. Analysts at the time triangulated — RDMA at 200 Gbps, per-die HBM figures — to a mix of Huawei Ascend 910C, Moore Threads, and MetaX accelerators, none of it confirmed.
For LongCat-2.5 the official reticence holds for training, but inference gives the game away. The SGLang-FluentLLM self-hosting reference is a 194-chip cluster of Huawei Ascend Atlas A2 NPUs: two for image preprocessing, 64 for prompt prefill, 128 for decoding. There is no Nvidia path. There is no CUDA kernel. The quantization scheme assumes an INT4 tensor core that Huawei has and Nvidia's consumer chips, pointedly, do not expose the same way. The hardware Meituan is paying attention to is the hardware it can buy.
A food-delivery company publishing a Huawei-native agent-serving guide alongside its frontier model is not neutral infrastructure work. It is a vote.
What this does to the export-control thesis
The US export-control story, through two administrations, has rested on a load-bearing assumption: China's frontier labs can shortcut a generation or two by buying or smuggling Nvidia hardware, but they cannot stand up a domestic-silicon training path for trillion-parameter-class models at throughput and cost that matter commercially. The June 2026 LongCat-2.0 release put a crack in that story. The November 2026 LongCat-2.5 release — vision added, agent benchmarks competitive, weights open, Huawei self-hosting path documented — widens it.
Three things worth being specific about:
- "Competitive with Claude Opus on several benchmarks" is still vendor-reported. DeepSeek V4 made the same claim in April; the Loop covered that one too. Six months later, the frontier labs are still at the frontier. LongCat-2.5 is not Opus 5.5. It is a very serious open-weights alternative at an aggressive price, which is a different and in some ways more disruptive thing.
- Open weights under MIT, with a trademark carve-out, is the standard China playbook now. The license grants redistribution, modification, fine-tuning. It does not grant use of Meituan's trademarks, which is standard. Western relabellers are free to do what they will.
- The cheapest chip in the room wins the agent workload. If Meituan's 194-Ascend serving cluster delivers Terminal-Bench results within a few points of Opus at a fraction of the token cost, every Chinese enterprise with a coding-agent budget has a decision to make about where their next inference bill goes. The answer will not be Oracle.
What to watch
- Whether independent benchmarks hold up. LMSYS and the SWE-Bench Verified maintainers will eventually run LongCat-2.5 against the field. The gap between Meituan's in-house tables and a third-party leaderboard is the number that matters. Expect it within three weeks.
- Whether any US customer tries to run it anyway. MIT-licensed weights are MIT-licensed weights. The question is which US enterprise is the first to admit, in a post-earnings call, that its internal coding agent runs on LongCat-2.5 behind a proxy — and whether a Treasury letter follows.
- Whether Nvidia's interim CEO Colette Kress mentions it on Tuesday. The Q3 earnings call is tomorrow, November 17. Kress has to project continuity to a room of analysts who read The Decoder. A single question about Meituan's chip bill of materials in Q&A is the one question she cannot rehearse.
- Whether this ships to Hugging Face's number-one-trending slot. It is already climbing. If it stays there into the IPO roadshow week Anthropic is in the middle of, one of the trillion-dollar range's bull-case paragraphs gets a little harder to read with a straight face.
The Loop's standing position on China-AI releases is that the numbers are vendor-reported until a third party confirms them. The timing, however, is not.
* * *
Thanks for reading. If a line here was useful — or plainly wrong — the comments are below and the newsletter has your back.
Elsewhere in this issue
3 moreLetters
Arguments, corrections, questions. Anonymous comments allowed; be kind, be specific.