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

Libraries

How to use unsloth/DeepSeek-V4-Flash-GGUF with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="unsloth/DeepSeek-V4-Flash-GGUF",
	filename="UD-IQ1_M/DeepSeek-V4-Flash-UD-IQ1_M-00001-of-00003.gguf",
)

llm.create_chat_completion(
	messages = "No input example has been defined for this model task."
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use unsloth/DeepSeek-V4-Flash-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf unsloth/DeepSeek-V4-Flash-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
llama cli -hf unsloth/DeepSeek-V4-Flash-GGUF:UD-Q4_K_XL

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf unsloth/DeepSeek-V4-Flash-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
llama cli -hf unsloth/DeepSeek-V4-Flash-GGUF:UD-Q4_K_XL

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf unsloth/DeepSeek-V4-Flash-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
./llama-cli -hf unsloth/DeepSeek-V4-Flash-GGUF:UD-Q4_K_XL

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf unsloth/DeepSeek-V4-Flash-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
./build/bin/llama-cli -hf unsloth/DeepSeek-V4-Flash-GGUF:UD-Q4_K_XL

Use Docker

docker model run hf.co/unsloth/DeepSeek-V4-Flash-GGUF:UD-Q4_K_XL

LM Studio
Jan
Ollama
How to use unsloth/DeepSeek-V4-Flash-GGUF with Ollama:
```
ollama run hf.co/unsloth/DeepSeek-V4-Flash-GGUF:UD-Q4_K_XL
```

Unsloth Studio

How to use unsloth/DeepSeek-V4-Flash-GGUF with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for unsloth/DeepSeek-V4-Flash-GGUF to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for unsloth/DeepSeek-V4-Flash-GGUF to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for unsloth/DeepSeek-V4-Flash-GGUF to start chatting

How to use unsloth/DeepSeek-V4-Flash-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/DeepSeek-V4-Flash-GGUF:UD-Q4_K_XL

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "unsloth/DeepSeek-V4-Flash-GGUF:UD-Q4_K_XL"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use unsloth/DeepSeek-V4-Flash-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/DeepSeek-V4-Flash-GGUF:UD-Q4_K_XL

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default unsloth/DeepSeek-V4-Flash-GGUF:UD-Q4_K_XL

Run Hermes

hermes

Atomic Chat new

OpenClaw new

How to use unsloth/DeepSeek-V4-Flash-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/DeepSeek-V4-Flash-GGUF:UD-Q4_K_XL

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "unsloth/DeepSeek-V4-Flash-GGUF:UD-Q4_K_XL" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Docker Model Runner
How to use unsloth/DeepSeek-V4-Flash-GGUF with Docker Model Runner:
```
docker model run hf.co/unsloth/DeepSeek-V4-Flash-GGUF:UD-Q4_K_XL
```

Lemonade

How to use unsloth/DeepSeek-V4-Flash-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull unsloth/DeepSeek-V4-Flash-GGUF:UD-Q4_K_XL

Run and chat with the model

lemonade run user.DeepSeek-V4-Flash-GGUF-UD-Q4_K_XL

List all available models

lemonade list

Quality loss in smaller 4 bit?

by neptronix - opened 4 days ago

Discussion

neptronix

4 days ago

Hey unsloth.
I was wondering - deepseek flash v4 model weights came in MXFP4 or some other quantized 4bit format, right?
For a quant like IQ4_NL, can we expect some added quality loss because we're not quantizing from 16bit, as with most models?

Have 128gb and trying to figure out how much extra VRAM i need to have a good time with DS4 flash

danielhanchen

Unsloth AI org 4 days ago

UD-Q8_K_XL is full precision but it's not technically 8-bit, but rather 4-bit. Yes there will be some quality loss with IQ4_NL and UD-Q4_K_XL but very very less.

idleDaemon

4 days ago

•

edited 4 days ago

I'm finding the IQ3_XXS quant to be quite useful, so far not a lot of issues. Sometimes it makes malformed tool calls but that may be related to my use of opencode, and telling the model to try again has succeeded every time so far.

Running with an RTX Pro 6000 Blackwell and 128GB of DDR5, with a 327680 context window I get around 20tg and 400pp starting out. The token generation stays around 17 into 200k+ context lengths, but prompt processing slows down and gets below 200 at ~100K tokens.

For reference, I'm running it with this command:

llama-server \
    --model <path>/DeepSeek-V4-Flash-UD-IQ3_XXS-00001-of-00004.gguf \
    --threads 24 \
    --temp 1.0 \
    --top-p 1.0 \
    --min-p 0.0 \
    --ctx-size 327680 \
    --alias "DeepSeek-V4-Flash" \
    --host 0.0.0.0 \
    --no-mmap \
    -np 1 \
    -ngl auto \
    -fa on\
    --cache-type-k q8_0 \
    --cache-type-v q8_0 \
    --jinja --reasoning on --reasoning-format deepseek

My system can handle the full 1M context length, but so much of the model is offloaded to CPU (so KV cache can stay resident in VRAM) that there's a big hit to processing speeds. It's still usable if you're willing to wait, but not great.

I was under the impression that the KV cache was supposed to be smaller, so perhaps there's still some tinkering to be done with llama.cpp.

neptronix

3 days ago

UD-Q8_K_XL is full precision but it's not technically 8-bit, but rather 4-bit. Yes there will be some quality loss with IQ4_NL and UD-Q4_K_XL but very very less.

Thanks pardner!

I'm finding the IQ3_XXS quant to be quite useful, so far not a lot of issues. Sometimes it makes malformed tool calls but that may be related to my use of opencode, and telling the model to try again has succeeded every time so far.

I'm surprised at that because the source material is 4 bit and you'd think the rounding errors could get harsh coming from 4 bit.

Running with an RTX Pro 6000 Blackwell and 128GB of DDR5....

Getting 20 tokens/sec on so much CPU offloading is impressive actually. Maybe you'd have some fun with a low 4 bit quant of Step 3.5 flash on this setup.
Adding a RTX PRO 5000 48gb for 144gb total might yield a good time with these quants.

I hear some performance aspects are not implemented in llama.cpp yet. But, flash is super cheap online. I'm hopeful that it will rip once the support is there!

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