inclusionAI has published Ling-3.1-flash on Hugging Face, a model with 560 billion parameters in total, of which 25 billion work on any given token (a word fragment). It is a mixture of experts: 512 routed experts plus one shared, eight active per token. Roughly 4.5 percent of the model runs at a time, which is how something this big can be served for less than its size suggests. The card also claims a one-million-token context window, several hundred thousand words by a common rule of thumb.
The card calls it a native hybrid reasoner without defining the term. Its example requests carry an enable_thinking switch, and a separate cookbook covers turning reasoning on and off.
On performance, the card names tests such as Terminal-Bench 4.0, SWE-bench Pro and FrontierSWE and calls the results strong. It prints no scores and no rival models. These are inclusionAI’s own claims.
Missing entirely: the licence, who inclusionAI is, and the download size. The sample server uses eight H200 GPUs, so this is no laptop model. Until the licence appears, nobody can say whether a company may build on it.
Based on the inclusionAI model card for Ling-3.1-flash on Hugging Face, 15 October 2026.
