Ling-3.1-flash
Ant Group released the weights of Ling-3.1-flash, a 560B-parameter model that activates about 25B per token and supports a 1M-token context window.
Ant Group published the weights so developers can run the model on their own infrastructure, with deployment support in SGLang and vLLM. It targets coding, agent workflows and long-document professional tasks that use external tools. A mixed-precision deployment approach cuts weight memory from about 534GB to 292GB, so a Hopper GPU setup needs four GPUs instead of eight.
- Date
- Friday 16 October 2026
- Lab
- Ant Group
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Total parameters | about 560B about 25B activated per token | company |
| Context window | up to 1M tokens | company |
| Weight memory with mixed precision | 534GB to 292GB Hopper GPUs needed fell from eight to four | company |
The article says the weights follow an earlier introduction of the model, so the original announcement date may be earlier. The license was not stated.
Sources
This record was checked against its sources on 16 October 2026. How we check