Request: DFlash draft checkpoint for Mistral-Medium-3.5-128B

[shelizi]

Please support Mistral-Medium-3.5-128B. It is a dense model, and this draft is very important.

[PlyMxt]

Why? Officially, there is already a drafter called EAGLE3. Few people use the model; apart from the official one, there is no provider on OpenRouter. They aren’t on Hugging Face at all. Maybe someone does need the model, but its usage since its release on OpenRouter hasn’t been growing—it’s declining.

Moreover, it’s a dense model. To train a good DFlash for it, you’d need to spend a huge amount of resources and compute on it—128 billion active parameters.

Meanwhile, the models requested by the community—GLM-5.1 (40B), MiniMax-M2.7 (10B), DeepSeek-V4 Pro (49B) and Flash (13B), Kimi K2.6 (32B)—have active parameters… simply an order of magnitude less.