Instructions to use srv-sngh/gemma-4-12B-coder-fable5-composer2.5-nvfp4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use srv-sngh/gemma-4-12B-coder-fable5-composer2.5-nvfp4 with MLX:

# Make sure mlx-lm is installed
# pip install --upgrade mlx-lm

# Generate text with mlx-lm
from mlx_lm import load, generate

model, tokenizer = load("srv-sngh/gemma-4-12B-coder-fable5-composer2.5-nvfp4")

prompt = "Write a story about Einstein"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True
)

text = generate(model, tokenizer, prompt=prompt, verbose=True)

Notebooks
Google Colab
Kaggle
Local Apps Settings
LM Studio

How to use srv-sngh/gemma-4-12B-coder-fable5-composer2.5-nvfp4 with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "srv-sngh/gemma-4-12B-coder-fable5-composer2.5-nvfp4"

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "srv-sngh/gemma-4-12B-coder-fable5-composer2.5-nvfp4"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use srv-sngh/gemma-4-12B-coder-fable5-composer2.5-nvfp4 with Hermes Agent:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "srv-sngh/gemma-4-12B-coder-fable5-composer2.5-nvfp4"

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 srv-sngh/gemma-4-12B-coder-fable5-composer2.5-nvfp4

Run Hermes

hermes

MLX LM

How to use srv-sngh/gemma-4-12B-coder-fable5-composer2.5-nvfp4 with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "srv-sngh/gemma-4-12B-coder-fable5-composer2.5-nvfp4"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "srv-sngh/gemma-4-12B-coder-fable5-composer2.5-nvfp4"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "srv-sngh/gemma-4-12B-coder-fable5-composer2.5-nvfp4",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Gemma-4-12B-Coder (fable5 x composer2.5) - native MLX (nvfp4) for Apple Silicon

A GGUF-free, native-MLX build of yuxinlu1/gemma-4-12B-coder-fable5-composer2.5, runnable on Apple Silicon via KrillLM's native Swift + MLX engine (no Python at inference).

Credit

The model is @yuxinlu1's fine-tune of google/gemma-4-12B-it - a Python/algorithmic coding model that reasons in Gemma's thinking channel before emitting a solution. All capability credit is theirs; please star the original repo. This repo only re-packages the weights to run natively on MLX.

What this is

The fine-tune is published as GGUF (lossy k-quants) and as NVFP4 safetensors (compressed-tensors). This build is converted GGUF-free from the NVFP4 side into MLX:

decompress NVFP4 (compressed-tensors) -> bf16 (pure numpy; byte-exact to the source's stored values, self-checked), then
requantize to MLX nvfp4 with attention o_proj and the vision/audio projectors protected at 8-bit (the proven mixed-precision recipe).

So it is lossless by construction relative to the NVFP4 source. Converter + recipe: KrillLM tools/ and docs/GEMMA4_12B_CODER_FINETUNE.md.

Use

brew tap srvsngh99/krillm && brew install krillm
KRILL_ENABLE_THINKING=1 krillm run gemma-4-12b-coder

KRILL_ENABLE_THINKING=1 opens the model's reasoning channel (this is a thinking fine-tune); without it, it answers without reasoning. ~6.7 GB, runs on a 24 GB Mac at ~25 tok/s decode, ~1.6 s cold load.

Benchmarks

Apple M4 Pro, 24 GB, macOS. KrillLM v0.8.0, this nvfp4 build. Reasoning on, greedy (temperature 0), single sample.

Standard EvalPlus harness (leaderboard-comparable):

metric	pass@1
HumanEval (base)	85.4%
HumanEval+ (base + extra tests)	83.5%

For reference, KrillLM's own (more lenient) extraction harness scored HumanEval 89.6% reasoning-on / 82.9% reasoning-off on the same problems; the EvalPlus numbers above are the comparable ones. These reflect this fine-tune's capability measured through KrillLM's MLX runtime, not a KrillLM-vs-other-engine claim.

Notes

De-refused (not safety-aligned - add your own guardrails) and English/Python-centric, per the upstream model. This is a weight-only conversion; behavior is the upstream fine-tune's.