Instructions to use okayuji/KeyVoice-Refiner-2B-v0.1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use okayuji/KeyVoice-Refiner-2B-v0.1 with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="okayuji/KeyVoice-Refiner-2B-v0.1",
	filename="refiner-2b-v0.1-Q4_K_M.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use okayuji/KeyVoice-Refiner-2B-v0.1 with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf okayuji/KeyVoice-Refiner-2B-v0.1:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf okayuji/KeyVoice-Refiner-2B-v0.1:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf okayuji/KeyVoice-Refiner-2B-v0.1:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf okayuji/KeyVoice-Refiner-2B-v0.1:Q4_K_M

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 okayuji/KeyVoice-Refiner-2B-v0.1:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf okayuji/KeyVoice-Refiner-2B-v0.1:Q4_K_M

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 okayuji/KeyVoice-Refiner-2B-v0.1:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf okayuji/KeyVoice-Refiner-2B-v0.1:Q4_K_M

Use Docker

docker model run hf.co/okayuji/KeyVoice-Refiner-2B-v0.1:Q4_K_M

LM Studio
Jan

vLLM

How to use okayuji/KeyVoice-Refiner-2B-v0.1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "okayuji/KeyVoice-Refiner-2B-v0.1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "okayuji/KeyVoice-Refiner-2B-v0.1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/okayuji/KeyVoice-Refiner-2B-v0.1:Q4_K_M

Ollama
How to use okayuji/KeyVoice-Refiner-2B-v0.1 with Ollama:
```
ollama run hf.co/okayuji/KeyVoice-Refiner-2B-v0.1:Q4_K_M
```

Unsloth Studio new

How to use okayuji/KeyVoice-Refiner-2B-v0.1 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 okayuji/KeyVoice-Refiner-2B-v0.1 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 okayuji/KeyVoice-Refiner-2B-v0.1 to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for okayuji/KeyVoice-Refiner-2B-v0.1 to start chatting

Pi new

How to use okayuji/KeyVoice-Refiner-2B-v0.1 with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf okayuji/KeyVoice-Refiner-2B-v0.1:Q4_K_M

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": "okayuji/KeyVoice-Refiner-2B-v0.1:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use okayuji/KeyVoice-Refiner-2B-v0.1 with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf okayuji/KeyVoice-Refiner-2B-v0.1:Q4_K_M

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 okayuji/KeyVoice-Refiner-2B-v0.1:Q4_K_M

Run Hermes

hermes

Docker Model Runner
How to use okayuji/KeyVoice-Refiner-2B-v0.1 with Docker Model Runner:
```
docker model run hf.co/okayuji/KeyVoice-Refiner-2B-v0.1:Q4_K_M
```

Lemonade

How to use okayuji/KeyVoice-Refiner-2B-v0.1 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull okayuji/KeyVoice-Refiner-2B-v0.1:Q4_K_M

Run and chat with the model

lemonade run user.KeyVoice-Refiner-2B-v0.1-Q4_K_M

List all available models

lemonade list

KeyVoice Refiner 2B (v0.1)

A lightweight text-refinement LLM fine-tuned for cleaning up speech-to-text output. Specialized for tasks like punctuation, polite-form conversion, multilingual translation, summarization, and filler-word removal — not a general-purpose chat model.

If you want to try this model embedded in a finished app right away, you can use KeyVoice (a macOS voice-input app).

Model Details

Base model: Qwen/Qwen3.5-2B
Teacher model: Qwen/Qwen3.5-9B (sequence-level distillation)
Training method: LoRA (rank=32) + distillation (trained on the teacher's refined outputs)
Architecture: qwen3_5 (Dense)
Parameters: 2.27B
Quantization: Q4_K_M GGUF (refiner-2b-v0.1-Q4_K_M.gguf, 1.3 GB)
License: Apache 2.0 (inherited)

Intended Use

Designed for

Punctuation and line-break formatting on transcribed speech
Polite-form / honorific conversion (e.g. Japanese desu/masu tone)
Multilingual translation (validated across 46 languages)
Summarization / compression of verbose text
Filler-word removal (e.g. "um", "uh", or Japanese fillers like "eeto", "ano")
Custom prompt rewriting

Not designed for

General-purpose chat (refinement-specialized; conversational ability is weak)
Code generation or mathematical reasoning
Creative or long-form writing (use the Qwen3.5-9B teacher directly instead)
Safety-critical domains such as medical, legal, or financial advice

Performance

Internal evaluation (Phase C data-augmentation set):

Inference speed: ~104 tok/s (Apple Silicon M4, llama.cpp Q4_K_M)
Refinement-task evaluation: 46 languages, 100% pass rate
Memory: ~2 GB at Q4_K_M

How to Use

llama.cpp / llama-cpp-python

huggingface-cli download okayuji/KeyVoice-Refiner-2B-v0.1 \
  refiner-2b-v0.1-Q4_K_M.gguf \
  --local-dir ./models

./llama-cli -m ./models/refiner-2b-v0.1-Q4_K_M.gguf -p "..."

Chat template

Uses the same chat template as the base model (Qwen3.5-2B):

<|im_start|>system
You are an expert at refining transcribed text.
<|im_end|>
<|im_start|>user
{user instruction}
<|im_end|>
<|im_start|>assistant

Training Data

The training data was generated internally by the teacher (Qwen3.5-9B) producing refined outputs for the following tasks:

Transcribed speech → punctuation and line-break formatting
Plain text → polite form (desu/masu register)
Source text → translation across 46 languages
Verbose text → summarization / compression
Spoken text → filler-word removal

The exact dataset is not published (small-scale LoRA distillation by a single developer).

Limitations and Bias

Inherits the limitations of the base model. Knowledge cutoff and any biases of Qwen3.5-2B carry over.
Optimized for refinement tasks, so general-knowledge and reasoning ability may be lower than the base model in some cases.
Small model (2B) — extremely complex instructions may be misinterpreted.
For Japanese honorifics, the model targets the polite (desu/masu) register; nuanced use of higher honorifics (sonkeigo / kenjogo — respectful and humble forms) is weaker than the 9B teacher.

Citation

@misc{keyvoice-refiner-2b-v0.1,
  author = {okayuji},
  title = {KeyVoice Refiner 2B (v0.1)},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/okayuji/KeyVoice-Refiner-2B-v0.1}}
}

Base model citation

@misc{qwen3.5,
  title = {Qwen3.5},
  author = {Qwen Team},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Qwen/Qwen3.5-2B}}
}

License

This model is released under the Apache License 2.0, inheriting the license of the base model Qwen/Qwen3.5-2B and the teacher model Qwen/Qwen3.5-9B.

See LICENSE for full text.

Downloads last month: 179

GGUF

Model size

2B params

Architecture

qwen35

Hardware compatibility

4-bit

Model tree for okayuji/KeyVoice-Refiner-2B-v0.1

Base model

Qwen/Qwen3.5-2B-Base

Finetuned

Qwen/Qwen3.5-2B

Finetuned

(177)

this model