Instructions to use patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-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 patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-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 patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF:Q8_0
# Run inference directly in the terminal:
llama cli -hf patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF:Q8_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF:Q8_0
# Run inference directly in the terminal:
llama cli -hf patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF:Q8_0

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF:Q8_0
# Run inference directly in the terminal:
./build/bin/llama cli -hf patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF:Q8_0

Use Docker

docker model run hf.co/patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF:Q8_0

Llama new
LM Studio
Jan

vLLM

How to use patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-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": "patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF:Q8_0

Ollama
How to use patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF with Ollama:
```
ollama run hf.co/patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF:Q8_0
```
Unsloth Desktop

How to use patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF:Q8_0

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": "patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF:Q8_0"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF with Docker Model Runner:
```
docker model run hf.co/patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF:Q8_0
```

Lemonade

How to use patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF:Q8_0

Run and chat with the model

lemonade run user.DeepSeek-V4.1-Flash-1017-REAP192-GGUF-Q8_0

List all available models

lemonade list

Hermes Agent

How to use patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-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 patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF:Q8_0

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 patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF:Q8_0

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF:Q8_0

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 "patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-GGUF:Q8_0" \
  --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"

Any benchmarks available how this performs on 16k context?

by mcr-ksh - opened 12 days ago

Discussion

mcr-ksh

12 days ago

•

edited 12 days ago

Any benchmarks available how this quant performs?

mcr-ksh changed discussion status to closed 12 days ago

mcr-ksh changed discussion title from Max KV-cache on Jetson Thor? to Any benchmarks available how this performs on 16k context? 12 days ago

patrickbdevaney

Owner 10 days ago

Any benchmarks available how this quant performs?

Hey, thanks for asking! I just ran a quick 16k test on my AGX Thor.

KV cache is basically a non-issue on Thor with this model. V4.1 compresses its attention cache and stores it in FP4, so the whole context state is around 2 GiB at 32k tokens. What actually limits you is the weights, and how fast the runtime gets through a long prompt.

Test: one 15.4k-token prompt, 256 tokens out, single request, greedy. It was in thinking mode, so the output is reasoning tokens.

	CUDA server	llama.cpp (this GGUF)
prompt processing (15.4k tokens)	628 tok/s (~25 s)	86 tok/s (~3 min)
generation at 16k	30 tok/s (with speculative decoding)	9.5 tok/s
same prompt sent again	59 ms (prefix cache)	–
free memory while serving	8 GiB (32k context limit)	27 GiB (16k context limit)

At short context llama.cpp generates about 12 tok/s, and the CUDA server 25 tok/s plain, or 25–40 with speculation depending on the prompt.

llama.cpp

Upstream doesn't support this yet, so use my fork: https://github.com/patrickbdevaney/llama.cpp/tree/deepseek-v4.1-reap-thor

thor/build.sh builds it, and thor/start.sh starts an OpenAI-compatible llama-server on port 8080.
The default context is 16k. CTX=32768 thor/start.sh works too; it passes the extra keys the GGUF needs above 16k.
Vision works through the mmproj in this repo.
DSPARK=1 turns on the bundled drafter for speculative decoding, giving about 17–21 tok/s at short context. With it on, the output isn't always token-for-token the same as without it.
You also need the two Engram row files on local NVMe (not a USB drive). The "Engram row files" section in thor/README.md explains how.

Native CUDA server (the fast path on Thor)

Code: https://github.com/patrickbdevaney/deepseek-v4.1-flash-cuda-server
Weights: https://huggingface.co/patrickbdevaney/DeepSeek-V4.1-Flash-1017-REAP192-Thor (needs the same Engram row files)

It's a standalone CUDA engine written specifically for Thor (sm_110), so no llama.cpp or vLLM underneath. Speculative decoding is on by default, using the model's own drafter, and greedy output still matches plain decoding exactly. ./start.sh serves an OpenAI-compatible API on port 8000 with a 32k context limit. Set SEQMAX for more; I've tested allocating a 1M-token context. It also does image input, JSON schema / grammar constrained output, and caches previous prompts so follow-ups in a conversation are near instant.

One warning for both: Thor's memory is shared between CPU and GPU, so don't load a second model next to it. That can hard-reboot the box, and both start scripts will refuse to launch if another one is running.

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