Instructions to use nathanw1014/dflash2-capgate-repro with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use nathanw1014/dflash2-capgate-repro 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 nathanw1014/dflash2-capgate-repro
# Run inference directly in the terminal:
llama cli -hf nathanw1014/dflash2-capgate-repro

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf nathanw1014/dflash2-capgate-repro
# Run inference directly in the terminal:
llama cli -hf nathanw1014/dflash2-capgate-repro

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 nathanw1014/dflash2-capgate-repro
# Run inference directly in the terminal:
./llama-cli -hf nathanw1014/dflash2-capgate-repro

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 nathanw1014/dflash2-capgate-repro
# Run inference directly in the terminal:
./build/bin/llama-cli -hf nathanw1014/dflash2-capgate-repro

Use Docker

docker model run hf.co/nathanw1014/dflash2-capgate-repro

LM Studio
Jan
Ollama
How to use nathanw1014/dflash2-capgate-repro with Ollama:
```
ollama run hf.co/nathanw1014/dflash2-capgate-repro
```

Unsloth Studio

How to use nathanw1014/dflash2-capgate-repro 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 nathanw1014/dflash2-capgate-repro 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 nathanw1014/dflash2-capgate-repro to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for nathanw1014/dflash2-capgate-repro to start chatting

How to use nathanw1014/dflash2-capgate-repro with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf nathanw1014/dflash2-capgate-repro

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": "nathanw1014/dflash2-capgate-repro"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use nathanw1014/dflash2-capgate-repro with Docker Model Runner:
```
docker model run hf.co/nathanw1014/dflash2-capgate-repro
```

Lemonade

How to use nathanw1014/dflash2-capgate-repro with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull nathanw1014/dflash2-capgate-repro

Run and chat with the model

lemonade run user.dflash2-capgate-repro-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use nathanw1014/dflash2-capgate-repro with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf nathanw1014/dflash2-capgate-repro

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 nathanw1014/dflash2-capgate-repro

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use nathanw1014/dflash2-capgate-repro with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf nathanw1014/dflash2-capgate-repro

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 "nathanw1014/dflash2-capgate-repro" \
  --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"

DFlash2 capability-gate repro (INTENTIONALLY BROKEN test fixture)

Do not use this for serving. It is a deliberately corrupted draft model that exists only to reproduce a llama.cpp speculative-decoding bug and verify its fix (llama.cpp PR #27342, DFlash2 support).

dflash-27b-q8-selector-key-only.gguf is a byte-identical copy of inco.ai's DFlash v1 draft model for Qwen3.8-27B (MIT) with one metadata key injected: dflash.selector_top_k = 16. No selector tensors were added, so the metadata claims a DFlash2 selector lattice that the decode graph never builds.

What it reproduces

On llama.cpp builds where the host arms the DFlash2 path from the dflash.selector_top_k metadata key alone, run this file as -md with --spec-type draft-dflash against a Qwen3.8-27B target:

without the capability gate: no crash and no diagnostic. The v1 graph writes hidden states into t_h_nextn, the host decodes them as candidate token ids, and target verification rejects every draft. Measured on gfx1151/RADV (Vulkan): 0/1556 drafts accepted, 4.77 t/s vs ~11.4 t/s with speculation disabled - a silent 2.4x slowdown with correct output.
with the capability gate: a startup warning ("selector metadata but its decode graph does not build the selector lattice") and a clean fallback to the DFlash v1 draft path (162/295 accepted, 25.6 t/s in the same setup).

Rebuild it yourself

The 30-line generator (plus a second script covering the DSV4-backbone load-refusal case) lives on the fork branch dflash2-capgate-repro; the fix itself is the dflash2-capgate branch, commit 0b0f35d0e ("spec : capability-gate the DFlash2 draft path").

Credit: the underlying draft model is by inco.ai (MIT). This repository only adds one metadata key to make the mismatch reproducible.

Downloads last month: -

GGUF

Model size

2B params

Architecture

dflash

Hardware compatibility

We're not able to determine the quantization variants.

View all variants

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support