Meituan launches trillion-parameter LongCat-2.5 trained entirely on Chinese chips
Meituan released LongCat-2.5 on November 1, 2026, a 1.6-trillion-parameter agentic model with a million-token context window.
· Last verified: 3 Nov 2026 (LongCat blog, Meituan Newsroom, Economic Observer)
Summary
- Meituan released LongCat-2.5 on November 1, 2026, a 1.6-trillion-parameter agentic model with a million-token context window.
- The company says training ran entirely on domestic Chinese AI accelerators with under 80GB memory per device.
- Promotional pricing of $0.30 per million input tokens undercuts US coding models for enterprise buyers.
The latest
A natively multimodal model built for long-horizon agent tasks, LongCat-2.5 carries 1.6 trillion total parameters with roughly 48 billion active per token, a one-million-token context window and outputs up to 128,000 tokens through its API. Meituan said training was completed entirely on domestic Chinese AI accelerators, without naming the chip or the cluster size.
Details
- The predecessor: LongCat-2.5 succeeds LongCat-2.0, released in June 2026, which Meituan described as the first trillion-parameter model trained and served fully on domestic Chinese compute. The company said the new run used accelerators with less than 80GB of memory per device, and did not disclose which silicon was involved.
- The benchmarks: According to the company's own table, LongCat-2.5 scored 88.4 on Terminal Bench v2.1 against 70.8 for version 2.0, 64.6 on DeepSWE v1.1 against 9.7, 82.1 on PaperBench against 45.2, and 91.9 on BrowseComp against 79.9. Vision2Web came in at 78.6.
- Testing conditions: Meituan listed thinking enabled, temperature 1.0 and top_p 0.95, with Terminal Bench v2.1 run through Claude Code at a 512,000-token maximum sequence length and DeepSWE v1.1 through mini-swe-agent at one million tokens, averaged over three runs.
- The caveat: The company said starred figures in its table come from public sources while unstarred numbers were self-run and indicative only. Meituan claimed its results sit between Claude-Opus-4.8 and Claude-Opus-5 on key benchmarks, and exceed Claude-Opus-5 on some evaluations.
- Training method: Long-context training ran through mid-training and supervised fine-tuning on sequences up to one million tokens across trillions of tokens. Meituan said single-device throughput in long-context fine-tuning rose 86% against a DDP baseline using two-way sharded FSDP.
- Anti-gaming steps: Agentic coding data was drawn from GitHub pull requests linked to issues and validated with fail-to-pass and pass-to-pass tests. To block reward hacking, the company said task environments had no network access, git history was stripped, and a judge language model audited trajectories.
- Inference gains: W4A4 quantization of expert weights and activations on native INT4 units delivered a 43.3% increase in prefill throughput and 42.77% in decode throughput against a W8A8 INT8 baseline, with time-to-first-token down 15.91% and GMM kernel speedup averaging 1.73 times.
- Access and pricing: The model is available through the LongCat chat interface and an API compatible with both OpenAI and Anthropic formats, working with Claude Code, Codex and Hermes. Limited-time pricing is $0.30 per million uncached input tokens, $0.006 cached, and $1.20 for outputs.
- Commercial use: Meituan's newsroom cited a brand channel analysis in which the model produced a channel profit statement, flagging 23.2% growth in payment value alongside a 1.7% drop in channel profit. The company built a merchant AI platform called Miaoshou on LongCat models.
- Corporate line: CEO Wang Xing told the second-quarter earnings call on August 28, 2026 that Meituan's AI models and products would serve its core business rather than becoming a commodity supplier: "We will not be a token factory."
Between the lines
The pricing and the hardware claim work together. At $0.30 per million input tokens with OpenAI- and Anthropic-compatible endpoints, switching costs for developers are low, and a model trained wholly on domestic accelerators is insulated from export-control exposure. Meituan's own benchmark caveat, that unstarred numbers are self-run and indicative, means the Claude comparison remains the company's claim rather than independently established.
What's next
GitHub and Hugging Face links are marked coming soon with no date and no announced license. Whether weights are released, and under what terms, will determine how far adoption spreads beyond Meituan's API.