DeepSeek V4 Flash Vision Exp

“I tried it (full precision, DeepSeek-V4-Flash-Vision-Exp) and it turned out DeepSeek's vision capability is very limited. It can't read normal letter text. It's a single-grid design with no tiling, and everything gets squashed into a 384-to”

“you CAN NOW RUN IT ON 6K's! I'm testing it right now and it's awesome. ~230tps during reasoning, ~350tps on code generation, LMCache for instant prefix cache and ~7K prefill”

“currently on 2x Spark you lose -20% raw perf”

“just switched to deepseek v4 flash vision exp since it was released earlier this week”

“In actual operation, it only took about 1 minute to complete the task, the efficiency is truly amazing”

“In my anecdotal testing this model seems less capable than for example Qwen 3.8 27B, which for me was a big step up in vision e.g. in its precision to come up with coordinates for bounding boxes on an image using a normalized 0-100 scale.”

“DeepSeek v4 Flash Vision Exp looks to have an incremental improvement in addition to supporting vision and for users with dual DGX Sparks it's still the best model available and really the only large model that can be run at the native quan”

“Works very well so far, including vision”

“better at talking”

“another gift !!! Best summer ever for local LLM”