Full 510GB DeepSeek-V4.1-Flash on a 64GB RAM + RTX 3060 PC: 7.4 tok/s and 98.3% agreement
Muted_Context_1999 · reddit · 2026-10-24
A developer ran the full 510GB DeepSeek-V4.1-Flash-1017 (original weights, not a quant) on a 5-year-old PC (Ryzen 9 3950X, 64GB DDR4, RTX 3060 12GB) using the PROMETHEVS engine, now released as v0.2 (Apache-2.0). The compress-then-correct method computes each token on a 2-bit copy of experts and recomputes uncertain tokens (flagged by the model's top-two margin) with FP4 weights rebuilt exactly from the 2-bit copy plus a correction plane, sharing expert reads via DeepSeek's 5-token draft head. Results: up to 7.4 tok/s warm (vs 2.14 for exact decoding), 98.3% agreement on 960 audited tokens, 2s follow-up starts. Needs Windows + one RTX 30/40 card and 1.1TB NVMe; benchmark-report form open for testers.
More from Infra
- Miles v0.1.3 cuts first RL weight update from 99.5s to 3.9s, shares GPUs across train/rollout — hsu_byron · 2026-10-24
- 98%+ Tensor Core utilization shown in LLM training run, drawing admiration — gharik · 2026-10-24
- We spent millions on AI apps with zero revenue — but our old GPUs are now worth more than we paid — stephen280up · 2026-10-24
- Polymarket puts 24% odds on AI data centers launching to space by end of next year — Polymarket · 2026-10-24
- Polymarket jokes wolves are entering Finnish villages after Google cleared forest for AI data centers — Polymarket · 2026-10-24
- llama.cpp PR drops kq_mask for QSA selection, saving gigabytes of VRAM on Qwen Flash Next — jacek2023 · 2026-10-24