Become a sponsor to Magnvs Ltd
Hi, my name is Ryan Flavell, a 27-year-old solo dev from Birmingham, UK.
PROMETHEVS is my own research, done on my everyday home PC: one RTX 3060 12 GB, 64 GB of RAM, two NVMe drives. No team, no lab, no other GPUs.
The idea is compress-then-correct. A small 2-bit pass does most of the work, and any word it’s unsure about gets rechecked against the original weights. On my machine that runs the full 510 GB DeepSeek-V4.1-Flash at up to 7.4 tok/s warm on a code prompt (5.8 cold), with 98.3% the same words as the full model across 960 audited words.
It’s still slow for heavy chat, and it’s only measured on my rig so far. That’s what I’m working on.
The engine is free and open (Apache-2.0). Sponsoring buys me more hours on it instead of other work (and the ability to potentially extend to other hardware/setups).
If you’ve got a Windows PC with 64 GB RAM, a 12 GB+ RTX 30/40 card and about 1.1 TB of NVMe, running the benchmark and posting your numbers helps just as much.
Thank you 😉
Meet the team
-
Ryan Flavell MagnvsSeptemRyan - Founder and Lead Dev
Featured work
-
Magnvs7/Promethevs
PROMETHEVS: run the full DeepSeek-V4.1-Flash-1017 (510 GB) on a gaming PC with 64 GB RAM and a 12 GB GPU. PROMETHEVS engine (compress-then-correct), OpenAI-compatible local server, report and logs.