Pinned
Introducing SuperDeepseek-V4-Flash🚀
Strongest model you can run on 128~256GB VRAM/Unified RAM.
> Abliterated for the freedom
> fixed broken weights from abliteration with multi AI agent swarm.
> enhanced overall intelligence
> 1M context, peak 129tok/s on 2xDGX Spark
> Mixed
SuperDeepseek-V4-Flash running on 2xDGX Spark at peak 122tok/s.
Worked based on @MiaAI_lab recipe to run with DFlash MTP.
4bit quant with less than ~1% quality loss
Also it is abliterated and quality fixed to enhance the overall performance.
This is just amazing.




