For those who still associate LLM training with enterprise budgets and unreachable GPU clusters, the project shared on GitHub by developer Nicolas (nickname SevereTilt) sounds like a wake-up call in reverse. Spending about $200 on vast.ai — renting both consumer GPUs and an H100 — he brought a 1.1-billion-parameter model from scratch to a working chatbot, then published code, weights, and a CPU-based demo.

The technical pipeline is straightforward but instructive. The architecture is inspired by Gemma3, with targeted simplifications: a reduced context of 4096 tokens (no sliding window attention), a 32k vocabulary trained with SentencePiece, and tweaked hyperparameters to hit the desired parameter count. Pretraining used the FineWeb-Edu dataset on 20 billion tokens, with a clever twist: data from before 2023, so the model could be quizzed on “future” events to test coherence. The chat phase came from LoRA fine-tuning on OpenHermes.

The real news isn’t the model’s quality — the author himself rates it below Gemma3 1B and NanoChat — but the cost and hardware trajectory. Exploratory runs used an RTX 3090 (185M parameters, 19 hours) and an RTX 5090 (500M, 17 hours). The jump to one billion parameters required an H100: 13 hours for 2 billion tokens, then 130 hours for the final 20 billion. LoRA fine-tuning ran on a modest RTX 3060 over 52 hours. The total bill: roughly $200 at February/March rates, likely higher today due to GPU demand.

The experiment doesn’t stop at the cloud. The developer compiled a llama.cpp fork to support the architectural differences, quantized the model to Q2_K GGUF format, and got it running on a WearOS watch at about 2 tokens per second. He also deployed a demo site on Google Cloud Platform using only CPU and GGUF files, to analyze the base model’s logprobs and offer chat. The upshot: LLM inference, even after third-party training, can live entirely on owned or low-power hardware.

This case, far from isolated, reshapes the incentives for anyone considering on-premise or edge deployment. If a single person with $200 and solid skills can build a language assistant that never leaves a trusted perimeter, the TCO equation for businesses shifts: a cloud provider is no longer mandatory for every iteration. Data sovereignty stops being a luxury and becomes a concrete design choice, even for small teams or air-gapped environments.

There are clear limits, of course. The model’s quality is modest, the 4096-token context is tight for many real-world uses, and training cost is sensitive to GPU market swings. But what this story signals is structural: the commoditization of training is moving value from raw compute power to the ability to orchestrate data, pipelines, and deployment. The gap between a personal experiment and a working minimum viable product has irreversibly shrunk.