Instructions to use yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-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 yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-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 yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF: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 yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF: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 yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF:Q4_K_M

Use Docker

docker model run hf.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF:Q4_K_M

LM Studio
Jan

vLLM

How to use yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF:Q4_K_M

Ollama
How to use yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF with Ollama:
```
ollama run hf.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF:Q4_K_M
```
Unsloth Desktop

How to use yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF:Q4_K_M

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": "yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF with Docker Model Runner:
```
docker model run hf.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF:Q4_K_M
```

Lemonade

How to use yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.gemma-4-12B-coder-fable5-composer2.5-v1-GGUF-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-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 yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF: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 yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF:Q4_K_M

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 "yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF:Q4_K_M" \
  --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"

C# & Java

#27

pinned

by cnsiva - opened Jun 16

Discussion

cnsiva

Jun 16

•

edited Jun 16

I’ve noticed that this model is being discussed quite a bit. I’m curious to know if it’s suitable for a large codebase that uses both C# and Java, and that also includes SQL.

yuxinlu1

Owner Jun 16

Thanks for trying it! Quick honest take:
This v1 was distilled mainly from Composer-2.5 + Fable-5 chain-of-thought, focused on Python, function-level
tasks. C#, Java and SQL ride on the base gemma-4-12B-it's knowledge — I didn't specialize them, but tuning
didn't regress general coding either, so do a quick A/B on your own snippets.
For a large codebase: it's a 12B trained on single-turn examples, so think of it as a focused pair-programmer
(write/explain/refactor a function, draft a SQL query). It has a 256K context so you can feed it the relevant
files — but it won't autonomously navigate a whole multi-language repo yet.
v2 (in progress) will add broader language coverage + terminal/agentic data — exactly the multi-step,
multi-file workflows a big codebase needs. Feedback like yours shapes it. 🙏

cnsiva

Jun 17

Thank you for the update. Support for Java, C#, and SQL is essential for us to use this model in our real-time production codebase.

bdutta

Jun 17

Would be very interested in Java, Golang support in v2, along with terminal/agentic performance improvements. I believe the base model is not terrible, as I read another post indicating that terminal operations are happening reasonably well after right Jinja template is selected. IIRC it was on little-coder (pi) based agentic coding harness discussion group. Your model landing page / model card is already pretty good (in fact one of the most objectively informative ones I've seen for a distill/finetune), but if you could also show/share how you did it, would be great.

yuxinlu1

Owner Jun 18

@bdutta Really appreciate that — glad the card reads honest rather than hype. 🙏

Good news on languages: v2's coding data already includes an open-domain multilingual set aimed at exactly the
underrepresented ones — Java, Go, C#, SQL (plus Rust, C++) — so it covers what both you and cnsiva need, on top of the
new terminal/agentic multi-step data that v1 was missing.

And your read on the base is right: the native Gemma 4 tool protocol works fine once the client actually loops it (the
jinja path) — v1's weakness isn't the format, it's that it's a single-turn fine-tune, so it doesn't drive the
multi-step loop on its own. v2 targets that directly.

On how I did it — happy to share. Short version: black-box distillation from Claude / Composer chain-of-thought, and
for the coding data I made it verifiable — generate a CoT solution, actually run it against the task's tests, and keep
only the samples that pass — then a QLoRA fine-tune on a single consumer GPU (RTX 5090), merge, and quantize to GGUF.
If folks want, I'll write up a proper methodology post with the details. Would that be useful?

yuxinlu1 pinned discussion Jun 18

johnlockejrr

Jun 18

@bdutta Really appreciate that — glad the card reads honest rather than hype. 🙏

Good news on languages: v2's coding data already includes an open-domain multilingual set aimed at exactly the
underrepresented ones — Java, Go, C#, SQL (plus Rust, C++) — so it covers what both you and cnsiva need, on top of the
new terminal/agentic multi-step data that v1 was missing.

And your read on the base is right: the native Gemma 4 tool protocol works fine once the client actually loops it (the
jinja path) — v1's weakness isn't the format, it's that it's a single-turn fine-tune, so it doesn't drive the
multi-step loop on its own. v2 targets that directly.

On how I did it — happy to share. Short version: black-box distillation from Claude / Composer chain-of-thought, and
for the coding data I made it verifiable — generate a CoT solution, actually run it against the task's tests, and keep
only the samples that pass — then a QLoRA fine-tune on a single consumer GPU (RTX 5090), merge, and quantize to GGUF.
If folks want, I'll write up a proper methodology post with the details. Would that be useful?

cnsiva

Jun 19

•

edited Jun 19

Awesome! The V2 is released. Thank you! @yuxinlu1
Can we expect MLX version too?

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment