Coding-Monkey-Gemma-GGUF

Coding Monkey Gemma

Many people praise Gemma 4 26B for it's coding capabilities and for good reason. However the model is simply too large to run comfortable without heavy quantization compromises on a consumer GPU and so I turned to the 12B varient only to find that it fails at even basic coding tasks because it doesn't understand how to properly utilize the correct tools from the agent harness. So this model is a fine tune of Gemma 4 12B with reliable tool calling. The results speak for themselves showing a 71% reliability on exact tool calling on a 102-prompt held-out set which means correct ChatML syntax and parameter placement. This is a huge improvement over the base model which only had a 26.5% reliability on exact tool calling. I also saw a 16% increase in the number of tool calls actually emitted by the model which means that the model is more likely to actually use the tools it has access to which is huge for coding tasks. This model is great for anyone who wants to do local coding but is tight on VRAM.

n = 102 held-out prompts, none ever seen during training. This set is harder than an earlier 67-item eval because it now includes multi-argument database, knowledge-graph and shell tasks that are matched exactly.

Training Details

For the training of this model I used the QLoRA method with a 16GB 5070 Ti. It 3 hours and 48 minutes to complete (480 steps) however previous protyping totaled 30 hours aggregate. The training loss went from 1.695 to 0.120 and the held-out eval loss was 0.206. I could've trained for much longer but I ran into overfitting issues on previous attempts so I decided that training on a small concentrated dataset was the best option. The training dataset consisted of 5,211 high-quality agentic/tool-calling examples, every one formatted with Gemma 4's own chat template. The glaiveai was the bulk of the data but I also included AgentInstruct examples because they had real world bash and terminal actions. I converted those raw ReAct trajectories (bash, SQL, web, and knowledge-graph actions) into real structured Gemma tool calls instead of plain text. I split it 98/2 into 5,107 train / 104 held-out val, and the eval set was built only from the val split so none of the 102 scored prompts were ever seen during training preventing contamination.

The breakdown by source:

That's ~5,161 actual tool-calling examples (~99% of the set) trained over 3 epochs with QLoRA r=16/alpha=32, LR 2e-4 cosine, and an effective batch size of 32.

Quantization

This model has fp16, Q8_0, Q6_K, Q5_K_M, and Q4_K_M quantization available in the gguf format. You can find them in this repository

Usage

llama.cpp

To use this model with llama.cpp you can use the ggufs like so

llama-server \
	-m gemma-coder-Q4_K_M.gguf
	-c 32768

Ollama

Ollama requires you to create a model file with each gguf so you could have something like this

FROM ./gemma-coder-Q4_K_M.gguf
PARAMETER num_ctx 32768

then you can run it like so with the Modelfile in the same directory as the gguf

ollama create coding-monkey -f Modelfile

Lorivo

Deploy this model with Lorivo for a production-ready, OpenAI-compatible API endpoint. Lorivo supports serving fine-tuned LoRA adapters without requiring you to merge the adapter into the base model or manually manage the serving infrastructure. Note: Lorivo currently supports the LoRA adapter format containing adapter_config.json and the associated .safetensors adapter weights. The merged model or GGUF files are not supported for this deployment method.

lorivo deploy ./Coding-Monkey-Gemma