Models

Tiles is designed for everyday productivity tasks. It combines a carefully tested model, inference parameters, and agent harness, so users do not have to manage model selection, tuning, or scaffolding themselves.

Tiles uses Google’s gemma4-12b as the default model for everyday tasks. We use Unsloth’s Dynamic 2.0 GGUF release of Gemma 4 12B and enable first-party Multi-Token Prediction (MTP) speculative decoding out of the box, reducing decoding latency while preserving the target model’s output quality.

Tiles uses llama.cpp for local inference across supported platforms, with a common inference stack built around GGUF models.

Tiles targets everyday personal machines with at least 16 GB of system memory for real productivity tasks.