Instructions to use Tinysoft/Cosyvoice2-0.5B-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Tinysoft/Cosyvoice2-0.5B-GGUF with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="Tinysoft/Cosyvoice2-0.5B-GGUF",
	filename="cosyvoice.gguf",
)

llm.create_chat_completion(
	messages = "\"The answer to the universe is 42\""
)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use Tinysoft/Cosyvoice2-0.5B-GGUF with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf Tinysoft/Cosyvoice2-0.5B-GGUF:F16
# Run inference directly in the terminal:
llama-cli -hf Tinysoft/Cosyvoice2-0.5B-GGUF:F16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf Tinysoft/Cosyvoice2-0.5B-GGUF:F16
# Run inference directly in the terminal:
llama-cli -hf Tinysoft/Cosyvoice2-0.5B-GGUF:F16

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 Tinysoft/Cosyvoice2-0.5B-GGUF:F16
# Run inference directly in the terminal:
./llama-cli -hf Tinysoft/Cosyvoice2-0.5B-GGUF:F16

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 Tinysoft/Cosyvoice2-0.5B-GGUF:F16
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Tinysoft/Cosyvoice2-0.5B-GGUF:F16

Use Docker

docker model run hf.co/Tinysoft/Cosyvoice2-0.5B-GGUF:F16

LM Studio
Jan
Ollama
How to use Tinysoft/Cosyvoice2-0.5B-GGUF with Ollama:
```
ollama run hf.co/Tinysoft/Cosyvoice2-0.5B-GGUF:F16
```

Unsloth Studio new

How to use Tinysoft/Cosyvoice2-0.5B-GGUF 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 Tinysoft/Cosyvoice2-0.5B-GGUF 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 Tinysoft/Cosyvoice2-0.5B-GGUF to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for Tinysoft/Cosyvoice2-0.5B-GGUF to start chatting

Docker Model Runner
How to use Tinysoft/Cosyvoice2-0.5B-GGUF with Docker Model Runner:
```
docker model run hf.co/Tinysoft/Cosyvoice2-0.5B-GGUF:F16
```

Lemonade

How to use Tinysoft/Cosyvoice2-0.5B-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Tinysoft/Cosyvoice2-0.5B-GGUF:F16

Run and chat with the model

lemonade run user.Cosyvoice2-0.5B-GGUF-F16

List all available models

lemonade list

Tinysoft commited on Jul 8, 2025

Commit

d3e9f67

verified ·

1 Parent(s): b5feafe

Update README.md

Browse files

Files changed (1) hide show

README.md +3 -0

README.md CHANGED Viewed

@@ -7,5 +7,8 @@ base_model:
 This only works with the token ID directly. The tokenizer is completely busted.
 CosyVoice also has a rich pre- and post- processing on top of the LLM step, so you can't do TTS out of the box with llamacpp.
 Nevertheless, the LLM step is the slowest, and switching from pytorch to llamacpp yields 10x perf gain.

 This only works with the token ID directly. The tokenizer is completely busted.
+This is also missing the bias tensor on the decoding head. I don't think there's a way to do it without forking llamacpp.
+I have not benchmarked the impact on quality due to the missing bias tensor.
 CosyVoice also has a rich pre- and post- processing on top of the LLM step, so you can't do TTS out of the box with llamacpp.
 Nevertheless, the LLM step is the slowest, and switching from pytorch to llamacpp yields 10x perf gain.