Measurement wanted · request DR-011 · opened 2026-10-31
DeepSeek-V4.1-Flash on RTX 5090: not measured yet
We don't have a DeepSeek-V4.1-Flash speed run on RTX 5090 yet, so there is no tokens-per-second number here. People asked for it, and this page is where the result will appear. Anyone with that hardware can run it with the free llm-speed CLI, and we'll post it as a community run.
The smallest quant is a 233 GB download and the RTX 5090 has 32 GB of VRAM, so most of the weights stay in system RAM or are read from disk. The CLI records your total system RAM, rounded to 8 GB, with the run; it doesn't measure memory use.
What's being requested
- Model
- DeepSeek-V4.1-Flash · unsloth/DeepSeek-V4.1-Flash-1017-GGUF
- Hardware
- RTX 5090
- Suggested quant
- UD-IQ1_S, about 233 GB download
Memory guidance comes from the publisher, not from us: see Unsloth's guide.
Runs we already have
No DeepSeek-V4.1-Flash runs on any hardware yet.
Run it yourself
Stock llama.cpp loads this model from release b11870 on (PR #31086 merged upstream). Use b12018 or later. Release page.
Start llama-server with the model (flags below are a starting point; adjust offload to your machine), then point llm-speed at it. Add --llama-args='...' with the flags you used so the run records them.
pipx install llm-speed # or: pipx upgrade llm-speed
# llama.cpp b12018 or later; N = the smallest value that fits your VRAM
llama-server -m /path/to/DeepSeek-V4.1-Flash-1017-UD-IQ1_S-00001-of-00005.gguf --port 8080 \
-ngl 999 --n-cpu-moe N -ot per_layer_token_embd.weight=CPU -lm mmap -lzm on -fa on
# then, in a second terminal:
LLAMA_CPP_SERVER_URL=http://127.0.0.1:8080 \
llm-speed bench --backend llama.cpp \
--model /path/to/DeepSeek-V4.1-Flash-1017-UD-IQ1_S-00001-of-00005.gguf \
--workload chat-shortThe upload is signed and public. How contributing works · what the CLI sends.
Get an alert
Email alerts for this page are switching on soon. Until then, follow @llmspeed on X, where new runs are posted.