Mini-AGI: One Developer Trains a Byte-Level Language Model Continuously on 8GB VRAM
A developer called volotat has published Mini-AGI, an open-source experiment in continual language model training on consumer hardware. The project, posted to GitHub and discussed on Hacker News, is explicit about its limitations: it is a toy-level model, not a frontier system, and its weights are not yet public. What makes it technically interesting is how it handles three constraints that usually require choosing one over the other.
Three design constraints, solved together
8 GB VRAM limit. Training needs gradients and optimiser state, which is roughly three times the weights again. Mini-AGI handles this by storing weights as ordinary files on disk and paging only the active working set onto the GPU. Parameter count is bounded by free disk space rather than VRAM, which makes the ceiling much higher than the card.
No catastrophic forgetting. Standard fine-tuning on a stream of new data overwrites earlier knowledge. Mini-AGI addresses this by making reading and training the same event: the model processes a stream of text one chunk at a time, takes a gradient step on each, and uses the same code path for generation. There is no separate fine-tuning regime and no frozen base.
No tokenizer. The model’s alphabet is all 256 byte values, so it can ingest any data type without a vocabulary mapping step. There is nothing to fit and no data type that requires a new token set.
The model also grows new capacity dynamically when it runs short during training and prunes what is not being used — architectural self-modification during a live run.
What the author says
volotat is direct about scope in the README: “This is a small experiment to show that continual learning from a single stream of data without catastrophic forgetting is possible. Furthermore it is possible on modest hardware.”
The model is described as a “small toy-level model.” Frontier-level capabilities are explicitly not the goal. The weights are not published because the model is still completing its first pass over the training corpus, which the author expects to take another couple of weeks at the current rate.
Why it gets attention
Most personal AI deployments today are frozen models someone else trained. Fine-tuning adds a user-specific layer but does not restart the learning process, and standard continual training causes the model to forget prior knowledge. Mini-AGI’s claim is that both problems can be avoided simultaneously on hardware a developer already owns.
The HN thread reached the front page, and a summary on ML Times described it as a “continual learning byte-level language model that can train from scratch on a single 8 GB VRAM GPU, dynamically adjusting its architecture.” Whether the architectural self-modification holds up at larger scale is not something the current experiment can answer. But the constrained-hardware angle — 8 GB VRAM laptop, no quantisation, full training — makes it reproducible for a wide audience without cloud budget.