Instructions to use unsloth/DeepSeek-V4-Flash-0731-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use unsloth/DeepSeek-V4-Flash-0731-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-0731-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
llama cli -hf unsloth/DeepSeek-V4-Flash-0731-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-0731-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
llama cli -hf unsloth/DeepSeek-V4-Flash-0731-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-0731-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
./llama-cli -hf unsloth/DeepSeek-V4-Flash-0731-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-0731-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
./build/bin/llama-cli -hf unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q4_K_XL

Use Docker

docker model run hf.co/unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q4_K_XL

LM Studio
Jan
Ollama
How to use unsloth/DeepSeek-V4-Flash-0731-GGUF with Ollama:
```
ollama run hf.co/unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q4_K_XL
```

Unsloth Studio

How to use unsloth/DeepSeek-V4-Flash-0731-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-0731-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-0731-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-0731-GGUF to start chatting

How to use unsloth/DeepSeek-V4-Flash-0731-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-0731-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-0731-GGUF:UD-Q4_K_XL"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use unsloth/DeepSeek-V4-Flash-0731-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-0731-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-0731-GGUF:UD-Q4_K_XL

Run Hermes

hermes

Atomic Chat new

OpenClaw new

How to use unsloth/DeepSeek-V4-Flash-0731-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-0731-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-0731-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-0731-GGUF with Docker Model Runner:
```
docker model run hf.co/unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q4_K_XL
```

Lemonade

How to use unsloth/DeepSeek-V4-Flash-0731-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q4_K_XL

Run and chat with the model

lemonade run user.DeepSeek-V4-Flash-0731-GGUF-UD-Q4_K_XL

List all available models

lemonade list

KLD Benchmarks + why Q8_K_XL vs MXFP4 naming

#11

by danielhanchen - opened about 2 hours ago

Discussion

danielhanchen

Unsloth AI org about 2 hours ago

Hey folks - for folks asking:

Why did we name the MXFP4 quant "Q8_K_XL" and not MXFP4
Why other folks' "lossless" quants are 155GB vs our Q8_K_XL at 162GB
Is FP8 and Q8_0 equivalent?

FP8 and Q8_0 are NOT bitwise equivalent - we checked the RMSE and it's not 0 at all. We did KLD on Q4_K_XL (the 155GB one other folks ship), and it's 96% top-1% agreement, so it's NOT lossless. llama.cpp does not have a native FP8 data-type, so BF16 is a must.

We could have named it MXFP4, but Q8_K_XL is the correct naming convention - we thought of BF16, but that's wrong. Also it's not all MXFP4, but MXFP4 + FP8

So when you see other people's quants at 155GB - this is not lossless and if they communicate that it's lossless - this is wrong. Converting FP8 to Q8_0 is not a lossless operation but lossy.

Full KLD table:

quant	size GB	mean KLD	KLD 99%	KLD 99.9%	PPL	top-1 %
UD-IQ1_S	82.5	0.66155	6.5134	11.3667	8.0932	72.30
UD-IQ1_M	86.9	0.58962	6.0813	10.4555	7.7384	74.08
UD-IQ2_XXS	90.9	0.48487	5.3472	9.7837	7.0909	76.60
UD-IQ2_M	90.9	0.48388	5.3461	9.7245	7.0894	76.56
UD-Q2_K_XL	96.8	0.40766	4.7839	9.0557	6.6782	78.57
UD-IQ3_XXS	104.2	0.30789	3.8147	7.7732	6.1972	81.93
UD-IQ3_S	116.1	0.26895	3.4845	7.3190	6.0266	83.06
UD-Q3_K_M	128.1	0.15734	2.1807	4.9571	5.7028	87.06
UD-Q3_K_XL	128.2	0.15751	2.1456	4.9708	5.7085	87.05
UD-IQ4_XS	136.7	0.11525	1.5929	3.9121	5.5522	88.79
UD-IQ4_NL	136.7	0.11525	1.5929	3.9121	5.5522	88.79
UD-Q4_K_XL	155.1	0.01324	0.1675	0.5512	5.3380	96.04
UD-Q8_K_XL	161.9	-0.00000	0.0000	0.0001	5.3322	100.00

Note UD-IQ3_XXS has been updated to reduce max KLD, and also some 3-bit quants are smaller in size now

perelmanych

about 2 hours ago

•

edited about 2 hours ago

Thanks for the clarification! In terms of activated parameters in Gbs what difference we are looking at between UD-Q4_K_XL and UD-Q8_K_XL? If it is +7Gb per token then probably I can live with 96% top1 agreement. It would be nice if you can state explicitly this parameter for all quantizations.

danielhanchen

Unsloth AI org about 2 hours ago

@perelmanych UD-Q4_K_XL and UD-Q8_K_XL are virtually equivalent, except UD-Q8_K_XL for attention, other non MoE layers are BF16, whilst UD-Q4_K_XL is in Q8_0

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