Instructions to use Infatoshi/phonon-refiner-1.7b-drafter with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Infatoshi/phonon-refiner-1.7b-drafter 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 Infatoshi/phonon-refiner-1.7b-drafter
# Run inference directly in the terminal:
llama cli -hf Infatoshi/phonon-refiner-1.7b-drafter

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Infatoshi/phonon-refiner-1.7b-drafter
# Run inference directly in the terminal:
llama cli -hf Infatoshi/phonon-refiner-1.7b-drafter

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 Infatoshi/phonon-refiner-1.7b-drafter
# Run inference directly in the terminal:
./build/bin/llama cli -hf Infatoshi/phonon-refiner-1.7b-drafter

Use Docker

docker model run hf.co/Infatoshi/phonon-refiner-1.7b-drafter

Llama new
LM Studio
Jan
Ollama
How to use Infatoshi/phonon-refiner-1.7b-drafter with Ollama:
```
ollama run hf.co/Infatoshi/phonon-refiner-1.7b-drafter
```
Unsloth Desktop

How to use Infatoshi/phonon-refiner-1.7b-drafter with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Infatoshi/phonon-refiner-1.7b-drafter

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": "Infatoshi/phonon-refiner-1.7b-drafter"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Infatoshi/phonon-refiner-1.7b-drafter with Docker Model Runner:
```
docker model run hf.co/Infatoshi/phonon-refiner-1.7b-drafter
```

Lemonade

How to use Infatoshi/phonon-refiner-1.7b-drafter with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Infatoshi/phonon-refiner-1.7b-drafter

Run and chat with the model

lemonade run user.phonon-refiner-1.7b-drafter-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use Infatoshi/phonon-refiner-1.7b-drafter with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Infatoshi/phonon-refiner-1.7b-drafter

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 Infatoshi/phonon-refiner-1.7b-drafter

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Infatoshi/phonon-refiner-1.7b-drafter with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Infatoshi/phonon-refiner-1.7b-drafter

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 "Infatoshi/phonon-refiner-1.7b-drafter" \
  --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"

Phonon refiner 1.7B: DFlash2 drafter

A small speculative-decoding drafter for Phonon refiner 1.7B, the transcript-correction model in Phonon. The drafter guesses the refiner's next 7 tokens in one pass, and the refiner checks all of them in one forward pass. With greedy verification the output is what the refiner produces alone, except where two tokens are within floating-point noise of each other, because the batched check rounds differently from one-token decoding. On Phonon's engine (M4 Max) 300 of 300 synthetic rows (two runs) and 579 of 580 real dictations came out identical to decoding without the drafter.

The same file ships inside the refiner repository, which is what the Phonon app downloads. This repository exists so the drafter has its own card; you do not need both.

File	Bytes	md5
`phonon-refiner-1.7b-dflash2-drafter.gguf` (Q8_0)	257,073,856	`e8cf3293bd35be5e3d573d5a43a9256f`
`LICENSE`, `NOTICE`

Architecture

DFlash2, Phonon's extension of DFlash block drafting:

3 transformer layers, FFN width 4096, hidden size 2048 (the refiner's), 32 attention heads of 64 with 8 KV heads, conditioned on the refiner's layer inputs from 5 of its 24 layers;
a 16k-token draft vocabulary drawn from the refiner's 128k: a token outside it is never drafted, which costs speed and never changes the output;
7 drafted tokens per step (a block of 8 with the last verified token).

Training

Every target is the refiner's own greedy output, so the drafter learns to predict this exact model. The inputs:

the refiner's 78,056 public training prompts (Parakeet transcripts of public YouTube audio, term-repair rows, hard negatives; see the refiner's card);
90,000 rows derived from ShareGPT conversations (ShareGPT_V3_unfiltered_cleaned_split, listed as Apache-2.0), turned into dictation-like text with synthetic recognition noise and vocabulary lines.

The drafter starts from a public drafter trained the same way against stock LFM3-1.7B and is fine-tuned for 1 epoch on these 167,056 rows. Training used SpecForge (MIT) draft modules and objective. No personal dictations and no evaluation sets were used; a leak check against every evaluation set found no overlap.

Speed

Tokens accepted per refiner forward pass (greedy, 7 drafted tokens, held-out public rows, CUDA, batch 1): 5.36 on 300 ShareGPT-derived rows and 6.12 on 300 public YouTube rows. The drafter it was fine-tuned from, aligned to stock LFM3-1.7B, gets 4.88 and 5.55 on the same rows with the same target, so alignment adds +0.48 [+0.41, +0.56] and +0.57 [+0.51, +0.63] (paired bootstrap 95%). Acceptance by draft position: 0.90, 0.84, 0.78, 0.72, 0.66, 0.61, 0.56.

On Phonon's engine (M4 Max, the refiner's Q4_K_S GGUF, prefix cache), 300 short synthetic requests (mean 15 output tokens) took 77 ms median and 112 ms at p90 with the drafter, against 110 / 145 ms without it (two runs each, spread under 1 ms on the mean), at 5.0 tokens per refiner pass. 580 real dictations (mean 81 output tokens) took 120-133 ms median and 373-393 ms at p90 at 6.0 tokens per pass (two runs). The stock-aligned drafter it started from got 5.3 tokens per pass and 401-404 ms at p90 on the same rows.

Running it

The GGUF layout is read by Phonon's native engine (pe, drafter kind dflash2n, source in the Phonon repository under engine/) and by llama.cpp with Phonon's patch (third_party/llama.cpp/phonon-engine.patch, --spec-type draft-dflash -md <this file> --spec-draft-n-max 7). Stock llama.cpp does not load it.

License

A derivative of LFM3-1.7B through the refiner it was trained on, distributed under the LFM Open License v1.0. Commercial use is licensed only to organizations under 10 million US dollars in annual revenue. See NOTICE.