midas-llm-research/qwen3-8b-dynadflash
1B • Updated • 51
None defined yet.
MIDAS LLM Research is a group of independent student researchers interested in inference, speculative decoding, and ML.
Our work so far consists of DynaDFlash, a model trained to build on top of the work of DFlash. It consists of the same backbone, except it dynamically computes a gate according to the context in which the drafter is being run in order to change how much of each hidden feature is involved in the projection.
The models are publicly available here!
We would like to thank @modal-labs for their generous research grant that allowed us to pursue this project. All models below were trained on Modal!