DeepSeek V4.1 Flash on 4x RTX 3090: 524k context at 30-40 tok/s decode
chota23 · reddit · 2026-11-04
A Reddit user ran DeepSeek-V4.1-Flash-1017 (abliterated GGUF, mix-256G) locally on 4x RTX 3090 (all PCIe x16) + 256GB DDR4 using the Maya engine at 524k context.
- Prefill: 400–600 tok/s; Decode: 30–40 tok/s (up to 60 tok/s with dspark)
- Q8 quantization is much slower: 6–8 tok/s decode, 65 tok/s prefill
- Compared to the 0731 version, V4.1 Flash loops less; but for agentic tasks the author still prefers the faster Qwen 3.8 Next Flash (Q8)
Useful first-hand benchmark data for anyone building a multi-GPU home rig for large open models; the author is soliciting speed comparisons from other engines (xolotl, strata).
More from Infra
- RTX 5090 Street Prices Top $9,000 as Nvidia GPU Costs Climb — teleforce · 2026-11-04
- Qdrant beats Meilisearch 2x faster and 9x less memory on production data — qdrant_engine · 2026-11-04
- Antseed hires BD lead to build a peer-to-peer open market for AI inference with 700+ models — ravikantagrawal · 2026-11-04
- Texas set to overtake Virginia as top US data center state, with most projects still early-stage — asusarla · 2026-11-04
- How one AI agent cut serving costs 37% on 1M weekly messages, in 13 production changes — AccBalanced · 2026-11-04
- One-line fp16 fix doubles Qwen-Image 2.1 speed on RTX 20-series in ComfyUI — xrailgun · 2026-11-04