inclusionAI releases Ling-3.1-flash, an MIT MoE model

inclusionAI has published Ling-3.1-flash-1019, a text-generation language model now available on Hugging Face. The release arrives under the permissive MIT license and is built on a hybrid mixture-of-experts (MoE) architecture, with weights distributed in the FP8 format.

The "flash" label positions this as a speed-oriented member of the Ling line. MoE designs activate only a subset of their parameters for any given token, which lets a model carry a large total capacity while keeping per-query compute lower than a comparable dense network. Pairing that with FP8 weights further trims the memory footprint, a combination aimed at teams that want to self-host without the overhead of a full-precision checkpoint.

Why it matters

Permissive licensing remains one of the biggest practical differentiators in open-weights AI, and the MIT terms here give developers room to adapt and deploy the model commercially. Key points from the release:

• Hybrid mixture-of-experts architecture for efficient inference
• FP8 weights to reduce memory requirements
• MIT license with few usage restrictions

inclusionAI has not published parameter counts, context length, or benchmark figures alongside this initial listing, so prospective users will want to validate performance against their own workloads. As the first entry in the Ling-3.1-flash family, it establishes a baseline the company can iterate on; the model card on Hugging Face is the authoritative place to track specs and any updates.