Tried running 748B DeepSeek-V4.1-Flash on a €2k PC
With the release of DeepSeek-V4.1-Flash-1017 I decided to try running it on my machine.
Proof1 community thread
Key points
- I bought a used PC with an RTX 3090 24 GB, 128 GB DDR4 and an Intel i9-9900X last year for under €2,000.
- It is of course the smallest quant available (1-bit), but I was still happy to see it working at 7.2 tok/s.
- Qwen models actually work pretty well on my setup.
- llama-server -m DeepSeek-V4.1-Flash-1017-UD-IQ1S-00001-of-00005.gguf \ -ngl 99 --n-cpu-moe 36 -ot "perlayertokenembd.weight=CPU" \ -lm mmap -lzm on -t 10 -c 16384 --jinja \ --temp 1.0 --top-p 1.0 --top-k 0 --min-p 0.0 \ --chat-template-kwargs '{"enablethinking":false}'
Sources (1)
- [1]Tried running 748B DeepSeek-V4.1-Flash on a €2k PCr/LocalLLaMA (top, daily) · Oct 21, 06:11 PM
With the release of DeepSeek-V4.1-Flash-1017 I decided to try running it on my machine.
I bought a used PC with an RTX 3090 24 GB, 128 GB DDR4 and an Intel i9-9900X last year for under €2,000.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 21, 2026Ran DeepSeek V4.1 Flash 1017 on a single RTX 5090
- Oct 21, 2026Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster
- Oct 20, 2026ornith-ai/Ornith-1.5-122B (Qwen finetune)
- Oct 20, 2026[AINews] DeepSeek DeepInfer: Open Source SOTA Inference running 155T toks/day/cluster making $22m/day at ~80% margin for a 81 day payback
- Oct 19, 2026unslothai/unsloth v0.1.909-beta: DeepSeek-V4.1 & FLUX-3
- Oct 19, 2026DeepSeek V4.1 Flash at 2.61 bpw + Engram: 26 t/s on dual RTX 3090s with dsx
