Instructions to use sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4 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 sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4
# Run inference directly in the terminal:
llama cli -hf sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4
# Run inference directly in the terminal:
llama cli -hf sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4

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 sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4
# Run inference directly in the terminal:
./build/bin/llama cli -hf sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4

Use Docker

docker model run hf.co/sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4

Llama new
LM Studio
Jan

vLLM

How to use sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4

Ollama
How to use sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4 with Ollama:
```
ollama run hf.co/sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4
```
Unsloth Desktop

How to use sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4 with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4

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": "sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4 with Docker Model Runner:
```
docker model run hf.co/sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4
```

Lemonade

How to use sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4

Run and chat with the model

lemonade run user.DeepSeek-v4-Flash-REAP-K150-NVFP4-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4 with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4

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 sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4 with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4

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 "sleepyeldrazi/DeepSeek-v4-Flash-REAP-K150-NVFP4" \
  --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"

sleepyeldrazi commited on Jun 23

Commit

f490ab9

verified ·

1 Parent(s): 6ebe1e1

Correct NVFP4 speed claims and document managed memory throughput cost

Browse files

Files changed (1) hide show

README.md +8 -6

README.md CHANGED Viewed

@@ -60,11 +60,12 @@ profile (273 GB/s LPDDR5X, ~97 GB/s managed-memory serving path):
 ### Why this mix?
 The bottleneck in MoE decode is gate+up expert projection (~2/3 of expert bytes).
-The donor GGUF uses IQ2_XXS here — but IQ2_XXS is **dequant-compute-bound** at
-~59 GB/s on GB10. NVFP4 reads more bytes (4.50 vs 2.06 bpw) but dequant is cheap
-(bandwidth-ish at ~140 GB/s), making it ~2.4× faster on the expert path. Down
-experts stay Q2_K (already saturating at ~160 GB/s). Attention and shared experts
-stay Q8_0 for quality.
 ### NVFP4: lossless MXFP4→NVFP4 conversion
@@ -107,7 +108,7 @@ make
 ### Inference (DGX Spark / NVIDIA GB10)
 ```bash
-# Managed memory required — K150 needs the single-residency path
 DS4_CUDA_MANAGED_MODEL=1 ./ds4 --cuda -m DeepSeek-V4-Flash-REAP-K150-hybrid.gguf --ctx 131072
 # With FP8-packed KV cache
@@ -132,6 +133,7 @@ DS4_CUDA_MANAGED_MODEL=1 ./ds4-server --cuda -m DeepSeek-V4-Flash-REAP-K150-hybr
 | 1M (FP32 KV) | ~121 GiB ⚠️ (borderline) |
 K150 needs managed memory (`DS4_CUDA_MANAGED_MODEL=1`) to avoid duplication.
 1M context requires FP8-packed KV (`DS4_KV_TURBO=1`).
 ## How It Was Built

 ### Why this mix?
 The bottleneck in MoE decode is gate+up expert projection (~2/3 of expert bytes).
+Standard GGUF quants use IQ2_XXS here. NVFP4 reads more bytes (4.50 vs 2.06 bpw)
+but offers substantially better precision at the same bit budget, with dequant
+bandwidth at ~140 GB/s on GB10. In practice, NVFP4 decode is slightly slower than
+IQ2_XXS due to the higher byte volume, but the precision gain is well worth the
+small throughput cost. Down experts stay Q2_K (already saturating at ~160 GB/s).
+Attention and shared experts stay Q8_0 for quality.
 ### NVFP4: lossless MXFP4→NVFP4 conversion
 ### Inference (DGX Spark / NVIDIA GB10)
 ```bash
+# Managed memory required — saves ~10 GiB RAM, costs 1-3 t/s
 DS4_CUDA_MANAGED_MODEL=1 ./ds4 --cuda -m DeepSeek-V4-Flash-REAP-K150-hybrid.gguf --ctx 131072
 # With FP8-packed KV cache
 | 1M (FP32 KV) | ~121 GiB ⚠️ (borderline) |
 K150 needs managed memory (`DS4_CUDA_MANAGED_MODEL=1`) to avoid duplication.
+This saves ~10 GiB of RAM at the cost of 1-3 t/s lower throughput.
 1M context requires FP8-packed KV (`DS4_KV_TURBO=1`).
 ## How It Was Built