Instructions to use deepseek-ai/DeepSeek-V4-Flash-Vision-Exp with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use deepseek-ai/DeepSeek-V4-Flash-Vision-Exp with Transformers:

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

pipe = pipeline("image-text-to-text", model="deepseek-ai/DeepSeek-V4-Flash-Vision-Exp")

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-V4-Flash-Vision-Exp")
model = AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-V4-Flash-Vision-Exp", device_map="auto")

Inference
HuggingChat
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use deepseek-ai/DeepSeek-V4-Flash-Vision-Exp with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp

SGLang

How to use deepseek-ai/DeepSeek-V4-Flash-Vision-Exp 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 "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp" \
    --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": "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp",
		"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 "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp" \
        --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": "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use deepseek-ai/DeepSeek-V4-Flash-Vision-Exp with Docker Model Runner:
```
docker model run hf.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp
```

DSpark acceptance rate seem seems quite a bit lower vs 0731

#11

by alexbi29 - opened 11 days ago

Discussion

alexbi29

11 days ago

66–71% versus 83–85% for 0731. Is this expected, seems like DSpark model was undertrained and/or is old?

BPAM

11 days ago

Per-workload datapoints from 2x DGX Spark (vLLM + DSpark, FP8 weights @ rev 86f746b) — the mixed-traffic 66–71% numbers hide a very wide per-content spread:

workload	accept @k=5	tok/step	accept @k=3	tok/step
count-to-300	0.974	5.88	0.997	4.00
code	0.464	3.31	0.642	2.93
prose	0.180	1.90	0.286	1.86

(prose accept nearly doubles at k=3, but tok/step barely moves, so throughput gains only ~9%.)

Two practical notes: (1) if you see 0.54 overall at k=5 on the DSpark recipe lineage, check the draft shared-expert loader fix ("Patch 4") first — its absence collapses the acceptance tail and mimics an undertrained drafter; (2) even with everything in place, code/prose acceptance stays well below the 0731 text model's, so ~12–17% lower tok/s on those workloads looks model-inherent. steps/s is flat across workloads at a given k (14.6–15.6 @k=5, ~16–17.3 @k=3) — the drafter, not step cost, is the lever.

alexbi29

10 days ago

I compared appearance rates on the same prompt 0731 vs new.

mwyborski

9 days ago

in sglang it got faster with only 1 nextn like 0731 instead of 3:
--json-model-override-args {"num_nextn_predict_layers":1}

Unkto

3 days ago

I was comparing
DeepSeek-V4-Flash-Vision-Exp-UD-Q8_K_XL.gguf + dspark-DeepSeek-V4-Flash-0731-BF16.gguf and
DeepSeek-V4-Flash-Vision-Exp-UD-Q8_K_XL.gguf + dspark-DeepSeek-V4-Flash-Vision-Exp-BF16.gguf

And dspark-DeepSeek-V4-Flash-0731-BF16.gguf delivers a higher draft acceptance rate than the native model across all tested query types, including "Describe what is shown in this photo" (sik).

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment