Meituan's LongCat team introduced LongCat-2.5 as a native multimodal model for long-horizon agentic tasks. The English technical blog is dated November 1, 2026; Meituan's Chinese newsroom post and secondary reports landed on November 2. The model has 1.6 trillion total parameters and activates about 48 billion parameters per token, building on LongCat-2.0 with stronger native vision and agentic coding.
Users can try the model at longcat.ai, and developers can call a newly available API. TechNode reports that GitHub and Hugging Face releases are planned to follow.
Key Facts
| Item | Detail |
|---|---|
| Model | LongCat-2.5 |
| Announced | Blog November 1, 2026; Meituan newsroom November 2, 2026 |
| Scale | 1.6T total parameters; ~48B activated per token |
| Focus | Native multimodality + long-horizon agent coding |
| Access now | Web demo at longcat.ai; API live |
| Weights | GitHub / Hugging Face "coming soon" (TechNode, November 2) |
| Training stack | Domestic AI accelerators end to end (Meituan newsroom) |
- Vision. Visual inputs are converted into multi-level discrete tokens so the backbone handles vision and text in one token format. The blog ties that design to visual coding, creative design, and 3D tasks.
- Agent RL. The team says it added more long-horizon trajectories to long-sequence training and scaled long-horizon reinforcement learning for end-to-end engineering on complex codebases and full-project builds. Failed real agent trajectories are turned into reproducible, verifiable training tasks.
- Context and infra. Mid-training uses sequences up to 1M tokens. Meituan says long-context SFT throughput on domestic accelerators rose 86% versus a DDP baseline in the measured setup. Inference uses W4A4 (expert weights and activations quantized to 4 bits). Online stress tests versus a W8A8 INT8 baseline show about +43% prefill and +43% decode throughput, with TTFT down about 16% and TPOT down about 20% (blog / Meituan newsroom).
- Benchmarks (vendor table). The blog's evaluation table compares LongCat-2.5 with GPT-5.5, GPT-5.6-Sol, Claude Opus 4.8, Claude Opus 5, and several open models. Meituan's Chinese newsroom summarizes that scores sit between Claude Opus 4.8 and Claude Opus 5 on many mainstream benchmarks, and exceed Opus 5 on some. Those are vendor-reported results; harness notes and asterisks for other models' public numbers appear in the blog footnotes.
Background
LongCat-2.0, released in June 2026 per Meituan's newsroom, was described as the team's first trillion-parameter model trained and served entirely on domestic accelerators. LongCat-2.5 keeps that "domestic model, domestic chips" line while adding native multimodality.
Meituan CEO Wang Xing, on the Q2 2026 earnings call (August 28), said the company would not become a "token factory" and that models should support Meituan's core local-commerce business. The newsroom ties LongCat-2.5 to merchant tooling such as the "Miaoshou" AI operations workbench already used by some chain brands. That product context is company messaging, not an independent adoption study.
Reactions and Impact
IT之家 and TechNode both repeated the 1.6T / ~48B activated figures, native multimodal design, and demo-plus-API availability on November 2. National Business Daily ran a short wire summarizing the same WeChat announcement. As of writing, we did not find independent third-party re-runs of the blog's full benchmark table beyond these reports.
What We Don't Know Yet
- Exact public release date and license for GitHub / Hugging Face weights
- Which domestic accelerator vendors and SKUs were used for the reported training and W4A4 numbers
- Independent verification of the vendor benchmark table under matched harnesses
- Pricing and rate limits for the new API beyond the public docs' media limits (images and video)