Full DeepSeek-V4.1-Flash-1017 (510 GB) on an RTX 3060 12 GB + 64 GB DDR4 + NVMe: 5.8 tok/s cold, 7.4 warm, >97% exact quality.

Hi all, I've been working on running the full DeepSeek-V4.1-Flash-1017 on an ordinary five-year-old gaming PC: Ryzen 9 3950X, 64 GB DDR4, an RTX 3060 with 12 GB, two NVMe drives, and about 1 TB of disk for the stores. I couldn't find a published figure for this model (or DS 4.1 original (or GLM)) on a 12 GB card with 64 GB of RAM, so as far as I know this is new at this hardware level. If I've missed one, please tell me and I'll compare. How it works: compress, then correct. Every token is computed on a 2-bit copy of the experts (0.61 of the bytes of the FP4 originals). The model's own margin between its top two candidates flags tokens it may have got wrong, and only those are recomputed with the original FP4 weights, which a stored correction plane rebuilds from the 2-bit copy. A draft head proposes blocks of five tokens so the expert reads are shared. Numbers (code prompt, 160 tokens, corrections on): 5.8 tok/s from a cold cache, 7.4 when the cache is warm from the same prompt, against 2.14 for plain exact decoding on the same machine. Reasoning and agent prompts run at 4.4 to 5.2. Quality, counted: on 960 audited tokens the 2-bit copy alone picks the same token as the exact path 97.4% of the time, and with the corrections 98.3%. A longer 1,101-token agent prompt is lower, 95.9% counting every difference. The audit runs at a margin gate of 0.9, so these count the differences it detected; tokens above the gate weren't checked. The reference is our engine's own exact path, which runs the original weights, not DeepSeek's reference implementation. That comparison is next. Why no code yet: this release is the report, the figures and every log behind the numbers. The engine's build scripts still use our drive layout and haven't been run on a clean machine, so I'd rather release it properly with a server than throw out something that doesn't run for you. It follows. Report and logs: github.com/Magnvs7/Promethevs I'd value feedback on the method and what you'd want to see tested next.