← sources

Deepseek V4.1-1017 - doing okay on RDNA2/3 + DDR4

source: reddit (reddit:LocalLLaMA)published 2026-10-23 03:56ingested 2026-10-24 02:40

retained text

Just a spec to performance report. Tried Maya today on the following:

- 3950x

- 96GB dual channel DDR4 at 3000mhz

- Rx 7900xtx 24GB + w6800 32GB (56GB together)

- pcie gen3 nvme drive

The Deepseek weights used were the mostly iq2 weights with Q6 ngrams. ~350GB on disk. Dspark left on.

I ended up getting wildly varying performance - less due to dspark hits and more due to when I/O was needed.

I got peaks as high as 19t/s, usually hovering around 13t/s, then dropping down to 3t/s for a few seconds, then right back up. **Average at the end of a long prompt was 9t/s decode. Prefill is ~150 t/s.** . Some prompts it rips through at 15 tokens a second, others it will crawl through at 3-5 tokens per second even if it's the same topic. Its very odd and I've been trying quite a few settings to figure it out.

The model itself is crazy smart - better than the iq4-xs quant of Qwen3.8-Flash-Next I've been running for certain - but it likes to think *A LOT* and with the already slow average of 9t/s decode I won't be using it for much outside of chatting even if it is the new best model on my machine.

document fetched 2026-10-24 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1xd3p2e/deepseek_v411017_doing_okay_on_rdna23_ddr4