Running quantized DeepSeek V4.1 on a single RTX 3090 at 20 tokens/s
GodComplecs · reddit · 2026-11-06
A Reddit user reports running a highly quantized DeepSeek V4.1 locally on a single RTX 3090 via Xolotl, set up with official parameters, achieving 20 tokens/s with 80GB DDR4 RAM and NVMe storage. Comparing answers with DeepSeek's official website, they find quality holds up well on high quantization settings — notably faster than the previous smaller V4 Flash, which ran at 8 tokens/s. A useful real-world data point for local deployment enthusiasts.
More from Infra
- LSEG cuts 5-year market backtest from hours to ~1 minute; Tampa General adds AI safety gate — NVIDIAAI · 2026-11-06
- TensorFold 1.2.3 ships faster Flash Next: +21% GB10 decode, M3 Ultra 64K-context reply in 8s — HankYeomans · 2026-11-06
- Autonomous Launches Local AI Workstations: 8-GPU Unit at $99,700 with Free 24-Hour Setup — dee_hw · 2026-11-06
- Polymarket odds put AI bubble burst by end of 2026 at just 8% on $2.6M volume — Polymarket · 2026-11-06
- KOSPI halted after 8.5% plunge in 3 minutes; SK Hynix -12%, Samsung -10% — ns123abc · 2026-11-06
- Rising energy and infrastructure costs may make AI pricier; efficiency is the smarter path — DavidLinthicum · 2026-11-06