Faster DeepSeek Flash on 2x3090 is here!
DeepSeek-Flash-1017:Maya-XS is almost identical in eval quality to the larger Maya-S quant, and runs much FASTER on a 2x3090+128gb desktop.
🔥 Decode: 32tps (up from 17tps on Maya-S)
🚀 Prefill: 1,400tps (up from 980 on Maya-S)
Thanks Project Maya!
github.com/mw00/project-maya
We have frontier-level intelligence at home now and I'm not even kidding.
DeepSeek-v4.1-Flash-1017:Maya-S, a 550B MOE model, running on a <$5k desktop PC from 2020.
🚀 Decode: 20tps
🔥 Prefill: 1,000tps
Thank you Project Maya!
github.com/mw00/project-maya
HW 2x3090+128gb+nvme
This is running on the same server I had set up for Qwen3.8-27b months ago
xitter.cf/axitek/status/21142481…
You can run a much SMARTER model on your hardware, with better results across the board.
Qwen3.8-27b vs DSF4.1-1017 (550b) - look at the eval difference between a 27b and a 550b model, running on the same 2x3090+128gb server.
Thanks Project Maya!
github.com/mw00/project-maya
You can run a much SMARTER model on your hardware, with better results across the board.
Qwen3.8-27b vs DSF4.1-1017 (550b) - look at the eval difference between a 27b and a 550b model, running on the same 2x3090+128gb server.
Thanks Project Maya!
github.com/mw00/project-maya
This would make a great Maya-S-uncensored quant wdyt @PeasantSmith
We're open-sourcing our OrcaCyber DeepSeek V4.1 Flash 1017 (Abliterated GGUF) weights — one of the strongest uncensored/abliterated DeepSeek V4.1 weights we've tested, with virtually no capability loss.
- 91.5% fewer refusals
- 92% MMLU vs. 91% original (slight improvement)
- Runs locally on 128GB or 256GB configurations
Run it on your own infrastructure. No cloud required.
Cyber Tier API: orcarouter.ai/models/orca/or…
Huggingface: huggingface.co/orcarouter/De…
It helps to have more VRAM - but don't discount RAM and NVME support.
Same model, DeepSeek-v4.1-Flash-1017:Maya-S on Maya:
🟠 Orange: 2x3090+128gb+nvme
🔵 Blue 7x3090+256gb+nvme
Speed difference is significant, but even 48gb vram is enough to run DSF.
github.com/mw00/project-maya
Axitek retweeted
DeepSeek-V4.1-Flash (256 experts, 274 GB) running at double-digit tokens/s on one RTX 5090 and 121 GB of RAM was unthinkable a year ago. Two engines do it today. We benched both - and the lesson is about RAM placement, not about anyone losing
A jailbreak with injected BASIC code is exactly my Saturday morning vibe
The ancient rule still stands: if it has a screen, it must run DOOM.
Yes, Switch 2 has a native version. I ran the original my way.
Just a warm-up. I've already gone much further.
Rip and tear.
#Switch2 #Jailbreak #Homebrew #DOOM #ns2jb