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.

ATH-MaaS logo

Can the Radeon RX 7800 XT run OvisOCR3?

OvisOCR3·853M·Radeon RX 7800 XT·16 GB VRAM

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

Yes. The Radeon RX 7800 XT runs OvisOCR3 fully in VRAM at FP16 / BF16 (about 2.2 GB of 16 GB).

best quant
FP16 / BF16
needs
2.2 GB
usable
16 GB

Estimates from the model's size.Try other contexts on the OvisOCR3 page →

Every quantisation of OvisOCR3 on the Radeon RX 7800 XT

QuantMemoryFit
Q4_K_M 1.1 GBfits
Q8_0 1.4 GBfits
FP16 / BF16 pick2.2 GBfits

Speed is a bandwidth-based estimate for generating text (624 GB/s); prompt processing and long contexts are slower.

Run OvisOCR3 on the Radeon RX 7800 XT

Using FP16 / BF16, the best fit for Radeon RX 7800 XT.

Ollama · F16 straight from Hugging Facedocs
ollama run hf.co/mradermacher/OvisOCR3-GGUF:F16
Ollama · Docker with ROCm (AMD)docs
docker run -d --name ollama --device /dev/kfd --device /dev/dri \
  -v ollama:/root/.ollama -p 11434:11434 ollama/ollama:rocm
docker exec -it ollama ollama run hf.co/mradermacher/OvisOCR3-GGUF:F16

The image bundles the ROCm runtime; the host only needs the amdgpu kernel driver.

llama.cpp · OpenAI-compatible server with a web UIdocs
llama-server -hf mradermacher/OvisOCR3-GGUF:F16

Needs a HIP (ROCm) build of llama.cpp. 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 \
  -v ~/.cache/llama.cpp:/root/.cache/llama.cpp \
  ghcr.io/ggml-org/llama.cpp:server-rocm \
  -hf mradermacher/OvisOCR3-GGUF:F16 --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/OvisOCR3-GGUF@F16

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

vLLM · Serve an OpenAI-compatible APIdocs
vllm serve ATH-MaaS/OvisOCR3

Needs vLLM built for ROCm; it supports Instinct MI200/MI300 and Radeon RX 7900 / 9000 cards.

vLLM · Smaller AWQ builddocs
vllm serve leleroyn/OvisOCR3-AWQ-W4A16
vLLM · Docker Compose (AMD ROCm)docs
services:
  vllm:
    image: rocm/vllm:latest
    command: vllm serve ATH-MaaS/OvisOCR3 --host 0.0.0.0 --port 8000
    ports:
      - "8000:8000"
    ipc: host
    devices:
      - /dev/kfd
      - /dev/dri
    group_add:
      - video
    security_opt:
      - seccomp=unconfined
    volumes:
      - ~/.cache/huggingface:/root/.cache/huggingface

AMD’s rocm/vllm image. API on port 8000.

SGLang · Serve an OpenAI-compatible APIdocs
sglang serve ATH-MaaS/OvisOCR3 --host 0.0.0.0 --port 30000

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

Transformers · Ask about an imagedocs
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="ATH-MaaS/OvisOCR3", dtype="auto", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
            {"type": "text", "text": "Describe this image."},
        ],
    }
]
out = pipe(text=messages, max_new_tokens=256)
print(out[0]["generated_text"][-1]["content"])

OCR models that fit the Radeon RX 7800 XT comfortably

All that fit →
ModelParamsBest quantMemoryLicence
PaddlePaddle logoPaddleOCR-VL-1.6PaddlePaddle
959MFP16 / BF16≈2.6 GBFits in memoryCommercial OK
PaddlePaddle logoPaddleOCR-VLPaddlePaddle
959MFP16 / BF16≈2.6 GBFits in memoryCommercial OK
tencent logoHunyuanOCRtencent
1.1BFP16 / BF16≈3.0 GBFits in memoryConditions
1BFP16 / BF16≈2.8 GBFits in memoryCommercial OK
853MFP16 / BF16≈2.4 GBFits in memoryCommercial OK
datalab-to logosurya-ocr-2.5datalab-to
689MFP16 / BF16≈2.0 GBFits in memoryConditions

Other GPUs that run OvisOCR3

Frequently asked questions

Can the Radeon RX 7800 XT run OvisOCR3?
Yes. The Radeon RX 7800 XT runs OvisOCR3 fully in VRAM at FP16 / BF16 (about 2.2 GB of 16 GB).
Which quantisation of OvisOCR3 should I use on the Radeon RX 7800 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.