Ran DeepSeek V4.1 Flash 1017 on a single RTX 5090
Been working over the last few days getting DeepSeek V4.1 Flash 1017 (Unsloth UD-IQ1S) running in Basalt (inference engine I'm developing).
Proof1 community thread
Key points
- Specs: RTX 5090 (400W) + RTX 5060 Ti 16 GB (180W), 128 GB DDR5.
- Baseline was llama.cpp PR #30817 with --fit, same hardware and quant.
- 5090 + 5060 Ti computing its own experts 63.5 tok/s 610 / 610 tok/s
- This was a personal experiment to see how far I could push it, there's probably some room to get around 80ish decode and ~1000 prefill, but on my hardware it doesn't make sense to keep going when I'm already getting over 300 tok/s on 3.8 FN.
Sources (1)
- [1]Ran DeepSeek V4.1 Flash 1017 on a single RTX 5090r/LocalLLaMA (top, daily) · Oct 21, 03:38 PM
Been working over the last few days getting DeepSeek V4.1 Flash 1017 (Unsloth UD-IQ1_S) running in Basalt (inference engine I'm developing).
Specs: RTX 5090 (400W) + RTX 5060 Ti 16 GB (180W), 128 GB DDR5.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 19, 2026unslothai/unsloth v0.1.909-beta: DeepSeek-V4.1 & FLUX-3
- Oct 19, 2026Qwen3.8-27B and Flash-Next on consumer hardware: engine numbers I could verify, and what they cost with 2026 RAM prices
- Oct 18, 2026Little Sweet Room Challenge | Strata Fork | Qwen 3.8 FN NVFP4 | DeepSeek Harness
- Oct 18, 202672 hours, four big open weights drops. The only one that fits in 24GB came from NVIDIA, of all companies.
- Oct 18, 2026You don’t need that expensive of a build for local llms.
- Oct 18, 2026Basalt: NVFP4 + faster prefill for Qwen3.8 Flash-Next on Blackwell (single 5090)
