Meituan — the company best known for delivering dinner across China — has released the open weights of LongCat-2.5, a 1.6-trillion-parameter multimodal model built specifically for agents that work through long coding tasks. The distinguishing detail is in the deployment guide rather than the benchmark table: as The Decoder reported, the self-hosting pipeline described in Meituan’s own documentation runs the model on Huawei Ascend accelerators, and Nvidia GPUs are never mentioned as an option anywhere in it.
🔍 THE BOTTOM LINE
A frontier-scale agent model you can download, modify, and self-host on entirely non-American hardware — under an MIT licence that permits commercial use — is the most direct challenge yet to both the closed-model business and the assumption that frontier AI is only buildable on Nvidia silicon. The weights dropped in the past week, and the Hugging Face checkpoint has already crossed 1,500 downloads.
What Meituan actually shipped
LongCat-2.5 is a mixture-of-experts model: 1.6 trillion total parameters with roughly 48 billion active per token. It handles text, images and video as input, toggles reasoning on and off, and is tuned for agents doing multi-step coding work — the same workload Western labs price at a premium. According to Meituan’s model card, the API accepts up to one million tokens of context and generates up to 128,000.
On agent benchmarks the model sits just below the current frontier rather than at it: Meituan lists 88.4 on Terminal Bench v2.1 against Claude Opus 5’s 89.1 and GPT-5.6 Sol’s 88.8, while trailing further behind on harder repository-bug-fixing tests. That is the same pattern China’s open-weight labs have shown all year — Kimi K3 cracked 2.8 trillion parameters and closed the visible benchmark gap, and GLM-5.3-Flash served 100 trillion daily tokens on domestic chips before its origin was even revealed. Per-token parity is arriving even where peak numbers still trail.
The part that is genuinely new: serving recipes, not just weights
Open-weight Chinese models are no longer news on their own. What changed this week is that Meituan published a complete self-hosting path for the model’s full-precision and 4-bit W4A4 variants on a cluster of 194 Huawei Ascend accelerators — 2 for image processing, 64 for prompt ingestion, 128 for output generation — with no guidance for Nvidia hardware at all. As The Decoder notes, the weights are quantised to 4-bit precision that Chinese chips process natively, lifting throughput by more than 40 percent in Meituan’s own load tests with accuracy held flat.
This continues a trajectory this site has tracked since June, when Huawei’s Ascend 910C fleet post-trained DeepSeek-V4-Pro’s 1.6 trillion parameters without a single Nvidia card — training, not just inference. Meituan’s predecessor model, LongCat-2.0, was trained on more than 50,000 Chinese chips. The new release extends the same playbook from “we can train on our chips” to “here is the operator’s manual for serving on them.” That second step is what turns a national capability into an exportable product: any data centre that can buy Ascend hardware can now buy a complete, licensed, frontier-adjacent agent stack to run on it. Meanwhile Washington keeps tightening its own perimeter — the DHS ordered DeepSeek stripped out of federal civilian systems in October — which does nothing to slow the hardware’s international spread.
Why a food-delivery company is in the frontier business
Meituan’s motive is not model sales. It runs one of the highest-volume dispatch workloads on earth — millions of riders, dynamic pricing, route agents — and agent-class models eat inference cost. Training its own model on its own chips, then giving the weights away under MIT, hedges the company against both Nvidia supply risk and Western API pricing. The official launch blog frames LongCat-2.5 as infrastructure for its own long-horizon reinforcement learning pipeline rather than a product.
For the wider market, the interesting number is the API price: roughly €0.27 per million input tokens at the promotional launch rate, an order of magnitude below frontier Western list prices. The model plugs into agent harnesses including Claude Code, Codex and Hermes through standard OpenAI- and Anthropic-compatible endpoints, so switching costs for a developer are near zero. China’s open-weight cohort has already put pressure on Western pricing from below; a MIT-licensed agent model with a published non-Nvidia serving recipe adds an infrastructure alternative to the price pressure.
FAQ
Can you actually run LongCat-2.5 without a Nvidia GPU? Yes — that is the point of the release. The published self-hosting guide describes a 194-node Huawei Ascend Atlas A2 cluster running the W4A4 quantised build, with quantisation done precisely because Chinese accelerators process 4-bit natively. Nvidia hardware is simply not addressed in the guide.
How good is it compared to Claude or GPT? Close on agentic coding benchmarks, behind on the hardest ones. The model posts 88.4 on Terminal Bench v2.1 versus 89.1 for Claude Opus 5, but trails most of the table on DeepSWE-style repository bug fixing, per The Decoder’s benchmark table.
Is it really free for commercial use? The weights are MIT-licensed, which permits commercial use, modification and redistribution. The hosted API is a separate product with promotional pricing reported at roughly €0.27 per million input tokens.
🔍 THE BOTTOM LINE
The benchmark gap between Chinese and American models has been narrowing all year; what LongCat-2.5 adds is the missing layer beneath the weights — a complete serving stack for domestic silicon, published in enough detail that anyone with a rack of Ascend cards can deploy a frontier-scale agent model. Export controls were written to slow exactly this convergence, and each release like this one shows how far behind the controls are running.
The weights released on the Hugging Face hub on 10 November; The Decoder’s full analysis followed on 15 November. Meituan’s Terminal Bench and DeepSWE figures are self-reported and have not been independently re-run as of publication.
