Instructions to use WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-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 WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-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 WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF
# Run inference directly in the terminal:
llama cli -hf WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF
# Run inference directly in the terminal:
llama cli -hf WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF

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 WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF
# Run inference directly in the terminal:
./llama-cli -hf WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF

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 WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF
# Run inference directly in the terminal:
./build/bin/llama-cli -hf WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF

Use Docker

docker model run hf.co/WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF

LM Studio
Jan
Ollama
How to use WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF with Ollama:
```
ollama run hf.co/WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF
```
Unsloth Desktop

How to use WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF

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": "WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF with Docker Model Runner:
```
docker model run hf.co/WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF
```

Lemonade

How to use WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF

Run and chat with the model

lemonade run user.DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-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 WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF

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 WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF

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 "WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B-exp-DS4-GGUF" \
  --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"

Experimental — most likely broken (renamed to -exp on 2026-08-12). This variant was produced by the second cascade step (REAM160 → REAM128 → REAM96). Severe multi-turn degradation was observed: generation typically collapses into gibberish from the second request onward. No quality evidence was ever collected for this release. Kept for archival reference only; do not use for inference. A rebuilt REAM96 (single-step prune from the base model, new calibration) will be published at the original repo id WaveCut/DeepSeek-V4-Flash-0731-REAM96-111B.

DeepSeek V4 Flash 0731 REAM96-111B DS4 Q2

This single-file DS4 GGUF was produced directly from the committed REAM96 native checkpoint. It preserves the 43-layer, 96-routed-expert, one-shared-expert, top-k 6 topology.

Runtime requirement: this is a DS4-specific GGUF. Use the patched WaveCut DS4 fork with dynamic expert-count support. Generic llama.cpp and Transformers loading are not claimed.

Quantization and identity

Routed gate/up tensors: IQ2_XXS.
Routed down tensors: Q2_K.
Attention projections, shared experts, and output: Q8.
MTP is absent.
DS4 revision: 6747e7718dd08f00b680d0c16231f2d59ec3747e.
Native NATIVE_READY.json SHA-256: 1a0aef0c8d4f736516ec1268b08d3ad698bef16f8e0ee90739fc0a0032831527.
Run: 20260811T072216Z.

Artifact	Bytes	SHA-256
`DeepSeek-V4-Flash-0731-REAM96-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix.gguf`	37,802,704,768	`e063760f1be5c5f30997461acc769d76e71d6c2f292645f22d7de3a5d769659c`
`ream96-imatrix.dat`	169,087,878	`897b8fdccbeb8d273752b07d5f11084857babe900fbbe7f183c3905e594a6db9`

The GGUF contains 1,328 tensors. Its quantization template is pinned to antirez/deepseek-v4-gguf@591165d69ad5d27cf826dcc2cf687549e966f893.

Native REAM and quantization methodology

The native checkpoint is a training-free cascade from the committed REAM128 release. It reduces every routed MoE layer from 128 to 96 experts using 3,072 calibration samples of 512 tokens each: 20% code, 10% math, 35% agent/tool, and 35% Russian/Cyrillic. Fresh layerwise saliency statistics, groups, merged experts, router mappings, and activations were computed for REAM96.

The importance matrix in this repository was generated fresh from the committed REAM96 native checkpoint; the REAM128 iMatrix was not reused. DS4 consumed it in strict iMatrix mode. Conversion was direct from Safetensors to GGUF without an intermediate F16 checkpoint.

Usage

Build or install the patched DS4 fork at the pinned revision, then run:

./ds4 \
  -m ./DeepSeek-V4-Flash-0731-REAM96-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix.gguf \
  --nothink

Remove --nothink only when intentionally enabling thinking with a suitable completion budget. Add backend, offload, context, and memory options for the installed CUDA or Metal build. The 37,802,704,768-byte model file alone is about 35.21 GiB; actual runtime memory also includes metadata, buffers, KV cache, and backend overhead.

Validation and integrity

Exact native lineage, GGUF byte size, tensor count, GGUF SHA-256, and fresh REAM96 iMatrix SHA-256 were verified and committed.
Comparative heldout PPL/NLL: not measured for this cascade.
Semantic continuation suite: not run for this cascade.
CUDA and Metal throughput: not benchmarked for this cascade.

The integrity checks prove artifact identity; they are not a quality or performance result.

Limitations

Do not reuse quality or performance measurements from the REAM128 or REAM160 releases.
Standard llama.cpp compatibility is not claimed; the 96-expert topology requires the patched DS4 runtime.
Memory and throughput depend on backend, offload, context length, batch size, and host I/O.