Pocket Mechanic: fp16 merged weights
google/gemma-4-E2B-it with the Pocket Mechanic LoRA merged in, saved as fp16 safetensors. This is the source the GGUFs are converted from.
For inference you almost certainly want the GGUFs: smaller and faster on CPU and Apple Silicon, and llama.cpp-ready. Use this fp16 version only if you need PyTorch / Transformers tooling specifically (e.g. further training, evaluation harnesses, hosted inference endpoints).
Quick load
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("MindFreakGamer/gemma-4-E2B-pocket-mechanic")
tokenizer = AutoTokenizer.from_pretrained("MindFreakGamer/gemma-4-E2B-pocket-mechanic")
Benchmark
Quantized to Q8_0 GGUF, this model reaches 81.3% of Claude Opus 4.7 teacher quality on a blind A/B judge (n=100, held-out cases). See the GGUF model card for the full benchmark table and how-to-use code.
License
Inherits the Gemma license.
Submission to the Hugging Face Build Small Hackathon, Backyard AI track (June 2026). Code: github.com/small-hack-huggingface/obd-intelligence.