On 2 November 2026, Meituan released LongCat-2.5, a natively multimodal model aimed at long-horizon agent work, according to its newsroom. Meituan says training ran entirely on domestic compute and that users can try the model at longcat.ai while developers can call a newly launched API. A technical post dated 2026-11-1 on the LongCat site describes the architecture, training stack, and benchmark table behind the release.
Key facts
| Item | Detail |
|---|---|
| Release | 2 Nov. 2026 (Meituan newsroom); technical blog 2026-11-1 |
| Scale | 1.6 trillion total parameters; ~48 billion activated per token (LongCat blog) |
| Focus | Long-horizon agentic tasks, complex coding, native vision |
| Access | Chat at longcat.ai; API launched same day (newsroom) |
| Inference | W4A4 on domestic accelerators; <80 GB memory per accelerator (blog) |
What Meituan released
Meituan’s newsroom frames LongCat-2.5 as the next step in its “国芯国模” line: a natively multimodal model for long agent tasks, built on LongCat-2.0. It says scores on several mainstream benchmarks sit between Claude-Opus-4.8 and Claude-Opus-5, with some results above Claude-Opus-5. The LongCat blog and homepage add stronger native visual understanding, agentic coding, and compatibility with Claude Code, Codex, Hermes, and other harnesses.
The numbers: 1.6T total, ~48B active
The LongCat post cites 1.6 trillion total parameters and roughly 48 billion activated per token. Meituan’s newsroom repeats the 1.6 trillion figure and says training used domestic compute throughout.
Built for long agent and coding tasks
LongCat-2.5 targets complex codebase development and full-project builds. The blog describes more long-horizon trajectories in long-sequence training, large-scale long-horizon RL across synthesized environments and mixed harnesses, and conversion of real agent failure trajectories into verifiable training tasks. Meituan’s newsroom cites mid-training sequences up to 1M tokens and multi-trillion-token cumulative training on domestic hardware.
Native multimodality via visual tokens
LongCat-2.5 maps visual signals to multi-level discrete tokens so one backbone processes images and text in a unified format, supporting visual coding, creative design, and 3D work per the blog.
Training and 4-bit inference on domestic accelerators
The blog optimizes long-context training and W4A4 inference on domestic AI accelerators with under 80 GB memory per accelerator. In Meituan’s own online stress tests versus a W8A8 INT8 baseline, it reports prefill throughput +43.3%, decode +42.77%, TTFT −15.91%, TPOT −19.72%, and 1.73× average GMM kernel speedup over W8A8.
Benchmarks Meituan reported
The LongCat evaluation table spans coding, general agent, and visual suites. Selected rows from Meituan’s published table (LongCat-2.5 vs. GPT-5.6-Sol and Claude-Opus-5; scores are Meituan-reported):
| Benchmark | LongCat-2.5 | GPT-5.6-Sol | Claude-Opus-5 |
|---|---|---|---|
| Terminal Bench v2.1 | 88.4 | 88.8* | 89.1* |
| BrowseComp | 91.9 | 90.4* | 90.8* |
| HLE w/tools | 63.1 | 58.0* | 64.7* |
| MCP Atlas | 83.1 | 81.8* | 85.8* |
| AutomationBench v1.0.6 | 51.8 | 45.8* | 50.3* |
| Vision2Web | 78.6 | 73.5* | 80.2* |
Competitor scores marked * come from public reports, per the blog footnote; other competitor figures were self-run by Meituan and are “indicative only.” We are not summarizing these rows as proof that LongCat-2.5 beats any single frontier model overall.
How to try it
Meituan’s newsroom says users can experience LongCat-2.5 on the official site and developers can invoke it through an API released the same day. The LongCat blog links chat and API access; GitHub and Hugging Face entries on that page are marked “Coming soon,” not available weights.
What’s not disclosed
The primary LongCat blog and Meituan newsroom release do not state API pricing, open-weight availability, GitHub or Hugging Face checkpoints for this version, or a public context-length limit for the product. Pricing was not given in the blog. Open-weight releases from other labs—such as Mistral Opens Up Large 4, Its 1-Trillion-Parameter Model or Moonshot's Kimi K3.1 Pro Is a 3T-Class Open-Weight Model—are separate stories; Meituan has not made comparable weight disclosures here.
What it means
Analysis (ours). The release pairs a frontier-scale agent model with Meituan’s merchant AI stack—newsroom examples of “妙手” and prior CEO remarks about supporting the core business rather than selling tokens alone. Domestic W4A4 serving may matter as much as benchmark rows for teams constrained to local accelerators.
What to watch
API pricing, any weight release, independent reruns of agent benchmarks, and production use through “妙手” and third-party harnesses beyond demos.