DeepSeek V4.1 Flash on a single RTX 3090 vs the API, side by side
A user compares DeepSeek V4.1 Flash 1017 running locally on one RTX 3090 with the Maya-S quant against the API on the same task: 25k tokens, low thinking effort and about an hour, slow but surprisingly good output.
Here's side by side comparison between DeepSeek-v4.1-flash running on my single RTX 3090 vs api. My system: - RTX 3090 24GB - 64GB DDR5 RAM - 1TB NVMe SSD Backend: Project-Maya Left one: DeepSeek-v4.1-flash-1017 from api - Total tokens: 60k - Thinking effort: High Right one: DeepSeek-v4.1-flash-1017: Maya-S quant - Total tokens: 25k - Thinking effort: Low - Total time: 1hr The one i'm running is slow, but it is really good. I'm suprised the result is this better. Like i say, smaller quant's of bigger models are not that bad.
This post quotes another post: “DeepSeek V4.1 Flash Maya-S quant runs one-shot on a 24GB RTX 3090”