TIM-360M: Transformer Inference Model

TIM (Transformer Inference Model) is a small language model (SLM) built from scratch by Tim. The project covers every stage: data sourcing, tokenisation, pretraining, post-training, and evaluation. Training and inference run on consumer grade hardware. No cloud training cluster is involved.

🤗 Download from HuggingFace 🦙 Run on Ollama

Why build a model from scratch?

Most practitioners fine-tune existing models. TIM goes further. It implements a modern decoder-only transformer from first principles, trains its own tokeniser, curates a multi-source corpus, and pretrains with a competitive small-model recipe, all within the limits of consumer grade GPUs. The result is a fully owned model family under the Apache 2.0 licence, where every design decision is documented and reproducible.

Why small models matter

Large models get the headlines, but small models do much of the real work. A model with a few hundred million parameters fits on a laptop, a phone or a single office server, and for many focused tasks that is all you need.

• Smart, efficient and focused. A small model tuned for a specific job can match a far larger general model on that job. Inference is rapid, even on edge devices and modest hardware.
• Private by design. Small models can be trained and fine-tuned on internal data. The entire workflow can be deployed entirely on premise. Your data never leaves your network, which matters in regulated industries such as energy, finance and government.
• Fast, local and yours. Responses arrive with low latency, there are no per-token API bills, and everything works offline. You hold the weights, so the model can never be withdrawn, altered or priced out from under you.
• Cheap to run. A small model draws a fraction of the power and cost of a large one. That makes it practical to serve many users, run continuously, or embed in products where a large model would never fit the budget.

Model at a glance

Trained on consumer-grade hardware

The whole pipeline runs on one Windows 11 workstation (training in WSL2):

These limits shaped every engineering decision: the 24 GB ceiling per GPU, the PCIe-only interconnect, and a hobbyist’s calendar. They led to a pretraining window of roughly two weeks.

Available models