https://unsloth.ai/docs/models/deepseek-v4.1-flash
Top Discussions
voogamer (⬆113): Holy crap, it’s just so… massive. MundanePercentage674 (OP) (⬆65): Keep in mind that its Engram Table Size alone is 196B–198B. ForsookComparison (⬆39): Keeping the “Flash” suffix implies the “Pro” version is still in the works. NoFunk (⬆23): Last night I downloaded a ds4-compatible version on my MacBook (https://huggingface.co/nazeshinjite/DeepSeek-V4.1-Flash-1017-ds4-GGUF), applied the mini ds4 patch with vision tower support, and hooked it up to run overnight on a Godot arcade game framework I wrote. It ran without a hitch—no errors, no infinite loops. It called tools whenever it was supposed to. It was pretty slow on the M5 Max (~15 t/s), but that was expected given the engrams and cold experts had to swap/fall back to the SSD. Running a 366GB model on 128GB of unified memory BeginningFew639 (⬆53): Only a 7GB difference? Seriously? Confident-Willow5457 (⬆45): That’s actually correct under Unsloth’s naming convention. The earlier DeepSeek V4 series was the same way. DeepSeek V4 itself is natively 4-bit, but Unsloth uses UD-Q8_K_XL here to indicate that all model weights remain at native/lossless precision, while UD-Q4_K_XL quantizes some tensors from BF16 down to Q8_0. tomz17 (⬆36): The dumbest naming scheme they could have backed themselves into… PermanentLiminality (⬆25): I need 0.2-bit.