Instructions to use AtomicChat/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 AtomicChat/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 AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16
# Run inference directly in the terminal:
llama cli -hf AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16
# Run inference directly in the terminal:
llama cli -hf AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16

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 AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16
# Run inference directly in the terminal:
./llama-cli -hf AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16

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 AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16
# Run inference directly in the terminal:
./build/bin/llama-cli -hf AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16

Use Docker

docker model run hf.co/AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16

LM Studio
Jan

vLLM

How to use AtomicChat/DeepSeek-V4-Flash-0731-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "AtomicChat/DeepSeek-V4-Flash-0731-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AtomicChat/DeepSeek-V4-Flash-0731-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16

Ollama
How to use AtomicChat/DeepSeek-V4-Flash-0731-GGUF with Ollama:
```
ollama run hf.co/AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16
```
Unsloth Desktop

How to use AtomicChat/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 AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use AtomicChat/DeepSeek-V4-Flash-0731-GGUF with Docker Model Runner:
```
docker model run hf.co/AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16
```

Lemonade

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

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16

Run and chat with the model

lemonade run user.DeepSeek-V4-Flash-0731-GGUF-BF16

List all available models

lemonade list

Hermes Agent

How to use AtomicChat/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 AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16

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 AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use AtomicChat/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 AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16

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 "AtomicChat/DeepSeek-V4-Flash-0731-GGUF:BF16" \
  --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"

AD-IQ3_XS

by djpista - opened 27 days ago

Discussion

djpista

27 days ago

Thank you for the upload.
I tried the AD-IQ3_XS quant, however I cannot seem to make it think on MAX with llama.cpp.
tried the official pi setup:
{
"id": "deepseek-v4-flash",
"name": "DeepSeek V4 Flash",
"contextWindow": 300000,
"maxTokens": 384000,
"input": ["text"],
"reasoning": true
"compat": {
"requiresReasoningContentOnAssistantMessages": true,
"thinkingFormat": "deepseek",
"reasoningEffortMap": {
"minimal": "high",
"low": "high",
"medium": "high",
"high": "high",
"xhigh": "max"
}
}

Also tried with --chat-template-kwargs '{"reasoning_effort":"max"}' or with --reasoning on
But with all of this just seems to think on low effort. (Either it doesn't think or very small like 1-2 lines and starts to write the file.)
None of this seemed to do anything.
On unsloth quant the chat template kwargs working.

My llama.cpp command:
./build/bin/llama-server --model ~/.cache/huggingface/hub/models--AtomicChat--DeepSeek-V4-Flash-0731-GGUF/snapshots/f4c7d285a98969f58aacfe0ebbf544def1a2f525/AD-IQ3_XS/DeepSeek-V4-Flash-0731-AD-IQ3_XS-00001-of-00004.gguf --ctx-size 384000 --load-mode none --no-warmup --temp 1.0 --top-p 1.0 --min-p 0.0 -np 1 -ngl 99 -fa on --alias "dsv4" --cache-ram 15360 --host 0.0.0.0 --port 9100 --jinja --chat-template-kwargs '{"reasoning_effort":"max"}'

Could you please give me some guidance on how to turn on max reasoning mode with these quants?

Thank you in advance.

worthant

Atomic Chat org 27 days ago

Hi @djpista ! Welcome :D
Thanks for the question!
I'll check and return with a proper guidance in a moment

worthant

Atomic Chat org 26 days ago

Hi again @djpista !
Thanks for finding that problem, it's on me. I've used generic chat template from late June instead of making/finding a new one and packing it into a GGUF quants. That's actually a severe problem and i need to update all quants.

The short reasoning you're seeing has a second cause from the same template - it gates reasoning retention on is_after_last_user, so in agentic loops every previous turn's resoning gets dropped once tool results come back. The current template keeps it.

Easy quick fix (NO redownload needed!):

I've just published the correct template at the repo root. download it;

curl -sLO https://huggingface.co/AtomicChat/DeepSeek-V4-Flash-0731-GGUF/resolve/main/chat_template.jinja

Then just add this temlate using those specific flags:

./build/bin/llama-server \
        --model ~/.cache/huggingface/hub/models--AtomicChat--DeepSeek-V4-Flash-0731-GGUF/snapshots/f4c7d285a98969f58aacfe0ebbf544def1a2f525/AD-IQ3_XS/DeepSeek-V4-Flash-0731-AD-IQ3_XS-00001-of-00004.gguf \
        --ctx-size 384000 --load-mode none --no-warmup \
        --temp 1.0 --top-p 1.0 --min-p 0.0 -np 1 -ngl 99 -fa on \
        --alias "dsv4" --cache-ram 15360 --host 0.0.0.0 --port 9100 \
        --jinja \
        --chat-template-file ./chat_template.jinja \
        --chat-template-kwargs '{"reasoning_effort":"max"}' \
        --reasoning-format deepseek

You don't need enable_thinking or --reasoning on. Thinking defaults to on and the server injects it into the template context. --reasoning-format deepseek is just explicit rather than relying on auto.

worthant

Atomic Chat org 26 days ago

Also, your provided model config has a JSON syntax error.

You're missing a comma after "reasoning":true
Fix:

{
  "id": "deepseek-v4-flash",
  "name": "DeepSeek V4 Flash",
  "contextWindow": 300000,
  "maxTokens": 384000,
  "input": [
    "text"
  ],
  "reasoning": true,
  "compat": {
    "requiresReasoningContentOnAssistantMessages": true,
    "thinkingFormat": "deepseek",
    "reasoningEffortMap": {
      "minimal": "high",
      "low": "high",
      "medium": "high",
      "high": "high",
      "xhigh": "max"
    }
  }
}

Note your map only emits "max" at the xhigh setting - every other level sends "high". Both are real levels in the 0731 template, but if you want max you have to actually select xhigh in the client.

requiresReasoningContentOnAssistantMessages: true is correct and now matters more than it did: the new template reads reasoning_content off prior assistant messages and preserves it through tool-call turns, so make sure your client is genuinely sending it back.

We'll insert the template into the GGUF metadata so the extra flag stops being necessary. Thanks again for catching this.

seleuco190ac

26 days ago

Are you planing to update the weights with the fix? I'll wait to download if it's the case.

worthant

Atomic Chat org 26 days ago

•

edited 26 days ago

Are you planing to update the weights with the fix? I'll wait to download if it's the case.

You dont need to update the weights just to use a chat template. Thats kind of a standard bug. Just pass it with the arguments i sent you above.

I will repack GGUFs, yes. But the chat template is stored in GGUFs metadata. Its simple to repack them, i just have a lot of tasks to do sorry. Will repack soon, for now pleass just use the jinja and chat template flags 😉

djpista

26 days ago

Hi again @djpista !
Thanks for finding that problem, it's on me. I've used generic chat template from late June instead of making/finding a new one and packing it into a GGUF quants. That's actually a severe problem and i need to update all quants.

The short reasoning you're seeing has a second cause from the same template - it gates reasoning retention on is_after_last_user, so in agentic loops every previous turn's resoning gets dropped once tool results come back. The current template keeps it.

Easy quick fix (NO redownload needed!):

I've just published the correct template at the repo root. download it;
curl -sLO https://huggingface.co/AtomicChat/DeepSeek-V4-Flash-0731-GGUF/resolve/main/chat_template.jinja
Then just add this temlate using those specific flags:
./build/bin/llama-server \
        --model ~/.cache/huggingface/hub/models--AtomicChat--DeepSeek-V4-Flash-0731-GGUF/snapshots/f4c7d285a98969f58aacfe0ebbf544def1a2f525/AD-IQ3_XS/DeepSeek-V4-Flash-0731-AD-IQ3_XS-00001-of-00004.gguf \
        --ctx-size 384000 --load-mode none --no-warmup \
        --temp 1.0 --top-p 1.0 --min-p 0.0 -np 1 -ngl 99 -fa on \
        --alias "dsv4" --cache-ram 15360 --host 0.0.0.0 --port 9100 \
        --jinja \
        --chat-template-file ./chat_template.jinja \
        --chat-template-kwargs '{"reasoning_effort":"max"}' \
        --reasoning-format deepseek
You don't need enable_thinking or --reasoning on. Thinking defaults to on and the server injects it into the template context. --reasoning-format deepseek is just explicit rather than relying on auto.

Thank you, it works great now.

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