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.

XHToken logo

Can the Radeon RX 6700 XT run Spark-X2.6-4B-Base?

Spark-X2.6-4B-Base·4.1B·Radeon RX 6700 XT·12 GB VRAM

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

Yes. The Radeon RX 6700 XT runs Spark-X2.6-4B-Base fully in VRAM at FP16 / BF16 (about 9.7 GB of 12 GB), at roughly 26 tokens/s.

best quant
FP16 / BF16
needs
9.7 GB
usable
12 GB
est. speed
~26 tok/s

Estimates for an 8K context; offload assumes 32 GB of system RAM.Try other contexts on the Spark-X2.6-4B-Base page →

Every quantisation of Spark-X2.6-4B-Base on the Radeon RX 6700 XT

QuantQualityMemoryFitSpeed
Q2_K Heavy loss3.4 GBfits~156 tok/s
Q3_K_M Noticeable loss3.8 GBfits~105 tok/s
Q4_K_M Small loss4.2 GBfits~85 tok/s
Q5_K_M Near-lossless4.5 GBfits~72 tok/s
Q6_K Near-lossless4.9 GBfits~62 tok/s
Q8_0 Lossless5.9 GBfits~48 tok/s
FP16 / BF16 pickLossless9.7 GBfits~26 tok/s

Speed is a bandwidth-based estimate for generating text (384 GB/s); prompt processing and long contexts are slower.

Measured speed on the Radeon RX 6700 XT

No community reports for this pair yet.

Report your own speed →

Run Spark-X2.6-4B-Base on the Radeon RX 6700 XT

Using Q6_K, the best fit for Radeon RX 6700 XT.

Ollama · Q6_K straight from Hugging Facedocs
ollama run hf.co/mradermacher/Spark-X2.6-4B-Base-i1-GGUF:Q6_K
Ollama · Docker with ROCm (AMD)docs
docker run -d --name ollama --device /dev/kfd --device /dev/dri \
  -e HSA_OVERRIDE_GFX_VERSION=10.3.0 \
  -v ollama:/root/.ollama -p 11434:11434 ollama/ollama:rocm
docker exec -it ollama ollama run hf.co/mradermacher/Spark-X2.6-4B-Base-i1-GGUF:Q6_K

ROCm doesn’t ship kernels for this card; -e HSA_OVERRIDE_GFX_VERSION=10.3.0 makes it run as the nearest supported target.

llama.cpp · OpenAI-compatible server with a web UIdocs
llama-server -hf mradermacher/Spark-X2.6-4B-Base-i1-GGUF:Q6_K

Needs a HIP (ROCm) build of llama.cpp. Start it with HSA_OVERRIDE_GFX_VERSION=10.3.0 in front for this card. Downloads the GGUF on first run.

llama.cpp · Docker with ROCm (AMD)docs
docker run -d --name llama --device /dev/kfd --device /dev/dri -p 8080:8080 \
  -e HSA_OVERRIDE_GFX_VERSION=10.3.0 \
  -v ~/.cache/llama.cpp:/root/.cache/llama.cpp \
  ghcr.io/ggml-org/llama.cpp:server-rocm \
  -hf mradermacher/Spark-X2.6-4B-Base-i1-GGUF:Q6_K --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/Spark-X2.6-4B-Base-i1-GGUF@Q6_K

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

SGLang · Serve an OpenAI-compatible APIdocs
sglang serve XHToken/Spark-X2.6-4B-Base --host 0.0.0.0 --port 30000

Needs SGLang’s ROCm build (Instinct MI300-class GPUs).

Transformers · Chat with the pipeline APIdocs
from transformers import pipeline

pipe = pipeline("text-generation", model="XHToken/Spark-X2.6-4B-Base", 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 Radeon RX 6700 XT comfortably

All that fit →
ModelParamsBest quantMemorySpeedLicence
4BQ8_0≈5.8 GBFits in memory~49 tok/sCommercial OK
3.2BQ8_0≈4.7 GBFits in memory~62 tok/sConditions
3.8BQ8_0≈5.5 GBFits in memory~52 tok/sCommercial OK
4.1BQ8_0≈5.9 GBFits in memory~48 tok/sCommercial OK
4BQ8_0≈5.8 GBFits in memory~49 tok/sCommercial OK
4.1BQ8_0≈5.9 GBFits in memory~48 tok/sCommercial OK

Other GPUs that run Spark-X2.6-4B-Base

Frequently asked questions

Can the Radeon RX 6700 XT run Spark-X2.6-4B-Base?
Yes. The Radeon RX 6700 XT runs Spark-X2.6-4B-Base fully in VRAM at FP16 / BF16 (about 9.7 GB of 12 GB), at roughly 26 tokens/s.
How fast does Spark-X2.6-4B-Base run on the Radeon RX 6700 XT?
An estimate from memory bandwidth gives about 26 tokens/s at FP16 / BF16.
Which quantisation of Spark-X2.6-4B-Base should I use on the Radeon RX 6700 XT?
FP16 / BF16: the largest one that fits. Expected quality: lossless 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.