Instructions to use cfigueiroa/Qwen3.8-27B-RVN-vision-MTP with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use cfigueiroa/Qwen3.8-27B-RVN-vision-MTP with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="cfigueiroa/Qwen3.8-27B-RVN-vision-MTP")

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("cfigueiroa/Qwen3.8-27B-RVN-vision-MTP", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use cfigueiroa/Qwen3.8-27B-RVN-vision-MTP with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "cfigueiroa/Qwen3.8-27B-RVN-vision-MTP"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cfigueiroa/Qwen3.8-27B-RVN-vision-MTP",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/cfigueiroa/Qwen3.8-27B-RVN-vision-MTP

SGLang

How to use cfigueiroa/Qwen3.8-27B-RVN-vision-MTP with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "cfigueiroa/Qwen3.8-27B-RVN-vision-MTP" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cfigueiroa/Qwen3.8-27B-RVN-vision-MTP",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "cfigueiroa/Qwen3.8-27B-RVN-vision-MTP" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cfigueiroa/Qwen3.8-27B-RVN-vision-MTP",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use cfigueiroa/Qwen3.8-27B-RVN-vision-MTP with Docker Model Runner:
```
docker model run hf.co/cfigueiroa/Qwen3.8-27B-RVN-vision-MTP
```

Qwen3.8-27B RVN + visão + MTP (embutido)

Cérebro RVN de 0bserverx (3 passadas ARA, o ficheiro RVN-Q4_K_M.gguf que eles publicaram sem NextN) + cabeça MTP/NextN enxertada (15/15 tensores blk.64.*, os mesmos que o Qwen original — a ARA não os altera) + projector de visão oficial.

Este combo (RVN 3-pass + NextN dentro + mmproj) não existe no repo do 0bserverx: eles converteram com --no-nextn.

Ficheiros

ficheiro	o que é
`RVN-Q4_K_M-MTP.gguf`	texto RVN Q4_K_M com `blk.64.nextn` 15/15 (~16.8 GB)
`mmproj-Qwen3.8-27B-Q8_0.gguf`	visão oficial `ggml-org`

Não uses -md / sidecar MTP. O acelerador já está no principal (~251 MiB, não 2.95 GB).

Correr

llama-server \
  -m RVN-Q4_K_M-MTP.gguf \
  --mmproj mmproj-Qwen3.8-27B-Q8_0.gguf \
  --jinja --reasoning off --device CUDA0

24 GB VRAM (RTX 3090 Ti): o Q4 + NextN embutido + mmproj Q8 cabe. Contexto longo come KV à parte.

Linhagem

Qwen/Qwen3.8-27B → trohrbaugh/Qwen3.8-27B-heretic-ara (1× ARA) → 0bserverx RVN (mais 2× ARA, --no-nextn) → este repo (NextN reposto a partir do dador Qwen/pass2 + mmproj oficial).

Apache-2.0. Não afiliado a Qwen, trohrbaugh ou 0bserverx.