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

Libraries

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

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

pipe = pipeline("text-generation", model="deepseek-ai/DeepSeek-V4-Flash")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-V4-Flash")
model = AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-V4-Flash")

Inference
HuggingChat
Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use deepseek-ai/DeepSeek-V4-Flash 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"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "deepseek-ai/DeepSeek-V4-Flash",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

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

SGLang

How to use deepseek-ai/DeepSeek-V4-Flash 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" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "deepseek-ai/DeepSeek-V4-Flash",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "deepseek-ai/DeepSeek-V4-Flash",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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

Claude 4.7 Reasoning into DeepSeek-V4-Flash for Open-Source Edge Intelligence

#26

by muneebkhanniazi - opened 4 days ago

Discussion

muneebkhanniazi

4 days ago

Claude 4.7 Reasoning into DeepSeek-V4-Flash for Open-Source Edge Intelligence

Project Objective
The goal of this project is to fine-tune the DeepSeek-V4-Flash model (284B MoE) using high-quality reasoning traces distilled from Claude 4.7 Opus. By training the model to utilize extended "thinking" chains, we aim to bridge the gap between frontier closed-source models and accessible, local open-source models that can run on consumer hardware.
Why This Matters (Impact)
Currently, high-level reasoning is locked behind expensive APIs. By distilling these logic patterns into DeepSeek-V4-Flash:

Accessibility: Users with 8GB–16GB VRAM can access Claude-level logic locally.

Performance: We expect to improve the model's score on SWE-bench and GSM8K by over 15% through explicit chain-of-thought training.

Efficiency: Utilizing TurboQuant (ICLR 2026), this model will support 1M+ context windows for local document analysis.

Dataset & Methodology

Dataset: We will utilize the reasoning-distill-opus-4-7-max-sft dataset (or similar community-leaked/generated traces) formatted for Supervised Fine-Tuning (SFT).

Technique: We will use Unsloth for 4-bit LoRA fine-tuning, specifically targeting the attention and MoE routing layers to adapt the model’s "thinking" behavior without degrading its base knowledge.

Compute Request

Requested Hardware: 1x NVIDIA H100 (80GB/94GB) for a duration of 48–72 hours.

Alternative: If a full grant is unavailable, we request access to a ZeroGPU Space with high-priority quota to host the training demo and evaluate the model in real-time.

Commitment to Open Source
Upon completion, the following will be released publicly on the Hugging Face Hub:

The fine-tuned GGUF and EXL2 weights (optimized for local use).

The full training script and adapter weights.

A comprehensive Model Card documenting the benchmarks and safety evaluations.

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