Example models

Our flagship demo. A 3B parameter model surgically reduced with Q4K quantization, then LoRA-finetuned on 271 joke examples. Tells coherent jokes, handles follow-ups, and answers general questions — all in a single 2 GB executable with a built-in web UI.

A task-specific English-to-French translator. Two layers pruned, Q4K quantization applied, then LoRA-distilled on 800 steps with a French translation dataset. 10/10 eval pass — accurate output with correct grammar and idiom.

Generates structured movie pitches from a genre and optional constraint. Built with ablation-guided layer selection — the first hoof model where compression decisions were driven by KL divergence scores rather than heuristics. 10/10 eval pass across 10 genres.

A focused coding assistant that writes, explains, and debugs code across Python, JavaScript, Rust, SQL, and Bash. SFT-finetuned on 367 curated examples. 9/10 on our eval suite — correct code, correct language, no padding.