selfhost.directory

The index of self-hostable projects — ranked, categorized and served with live updates.
Own your data and stop giving it to big tech.

xtools-at logo

Can the Arc A580 run DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic?

DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic·9.4B·Arc A580·8 GB VRAM

Saved in this browser, used on every model page.
Yes

Yes. The Arc A580 runs DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic fully in VRAM at Q4_K_M (about 7.2 GB of 8.0 GB), at roughly 49 tokens/s.

best quant
Q4_K_M
needs
7.2 GB
usable
8.0 GB
est. speed
~49 tok/s

Estimates for an 8K context; offload assumes 32 GB of system RAM.Try other contexts on the DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic page →

Every quantisation of DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic on the Arc A580

QuantQualityMemoryFitSpeed
Q2_K Heavy loss5.3 GBfits~91 tok/s
Q3_K_M Noticeable loss6.1 GBfits~61 tok/s
Q4_K_M pickSmall loss7.2 GBfits~49 tok/s
Q5_K_M Near-lossless8.0 GBtight~42 tok/s
Q6_K Near-lossless8.9 GBoffload~21 tok/s
Q8_0 Lossless11 GBoffload~9.1 tok/s
FP16 / BF16 Lossless20 GBoffload~2.7 tok/s

Speed is a bandwidth-based estimate for generating text (512 GB/s); prompt processing and long contexts are slower.

Measured speed on the Arc A580

No community reports for this pair yet.

Report your own speed →

Run DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic on the Arc A580

Using Q4_K_M, the best fit for Arc A580.

Ollama · Q4_K_M straight from Hugging Facedocs
ollama run hf.co/mradermacher/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic-GGUF:Q4_K_M

Ollama’s ROCm/CUDA builds don’t cover this GPU. Recent Ollama releases use Vulkan when its Vulkan backend and driver are installed (older ones needed OLLAMA_VULKAN=1); llama.cpp or LM Studio with Vulkan are the safer bet.

llama.cpp · OpenAI-compatible server with a web UIdocs
llama-server -hf mradermacher/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic-GGUF:Q4_K_M

Needs a Vulkan build of llama.cpp. Downloads the GGUF on first run.

llama.cpp · Docker with Vulkandocs
docker run -d --name llama --device /dev/dri -p 8080:8080 \
  -v ~/.cache/llama.cpp:/root/.cache/llama.cpp \
  ghcr.io/ggml-org/llama.cpp:server-vulkan \
  -hf mradermacher/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic-GGUF:Q4_K_M --host 0.0.0.0 --port 8080 -ngl 99

Server and web UI on port 8080.

LM Studio · Download with the lms CLIdocs
lms get mradermacher/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic-GGUF@Q4_K_M

Then load it from the LM Studio app or with `lms load`.

Transformers · Chat with the pipeline APIdocs
from transformers import pipeline

pipe = pipeline("text-generation", model="xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic", dtype="auto", device_map="auto")
messages = [{"role": "user", "content": "Explain self-hosting in one paragraph."}]
out = pipe(messages, max_new_tokens=512)
print(out[0]["generated_text"][-1]["content"])

Chat models that fit the Arc A580 comfortably

All that fit →
ModelParamsBest quantMemorySpeedLicence
8.2BQ5_K_M≈7.3 GBFits in memory~48 tok/sCommercial OK
8BQ5_K_M≈7.1 GBFits in memory~49 tok/sConditions
7.6BQ6_K≈7.0 GBFits in memory~45 tok/sCommercial OK
8BQ5_K_M≈7.1 GBFits in memory~49 tok/sConditions
ornith-ai logoOrnith-1.5-9Bornith-ai
9.7BQ4_K_M≈7.3 GBFits in memory~48 tok/sCommercial OK
ibm-granite logogranite-4.2-8bibm-granite
8.8BQ4_K_M≈7.0 GBFits in memory~53 tok/sCommercial OK

Other GPUs that run DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic

Frequently asked questions

Can the Arc A580 run DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic?
Yes. The Arc A580 runs DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic fully in VRAM at Q4_K_M (about 7.2 GB of 8.0 GB), at roughly 49 tokens/s.
How fast does DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic run on the Arc A580?
An estimate from memory bandwidth gives about 49 tokens/s at Q4_K_M.
Which quantisation of DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic should I use on the Arc A580?
Q4_K_M: the largest one that fits. Expected quality: small loss compared with the full-precision weights.

Memory and speed are estimates from the model's size and the card's memory bandwidth; real numbers depend on the runtime, drivers and context length. Community figures are medians of user reports.