G1-nano-base

A 60M-parameter GPT trained completely from scratch on a single 8GB-RAM NVIDIA Jetson device, without cloud infrastructure or multi-GPU setups. Base pretrained checkpoint with a native 2048-token context for raw text completion.

Overview

G1-nano-base is a 60.0M-parameter decoder-only causal language model trained from scratch under a single 8GB-RAM NVIDIA Jetson.

The project focuses on making a complete model training workflow — tokenizer, pretraining, fine-tuning infrastructure and export — work on modest hardware.

This is the base checkpoint. It predicts the next token and completes text; it is not a chat model and should not be expected to follow instructions.

Model variants

The instruction-tuned version of the same model is available as azerdsq/g1-nano-instruct.

What this version adds

Compared with the previous G0 Nano design, G1 Nano prioritizes a 2x native context length at approximately the same model size and training constraint. This base checkpoint does not include supervised instruction fine-tuning or a chat format.

Architecture

Llama-style decoder-only Transformer:

• Parameters: 60.0M
• Layers: 14
• Hidden size: 576
• Attention: Grouped-Query Attention, 9 query heads / 1 key-value head, head dimension 64
• Position encoding: RoPE, θ=10000
• Feed-forward network: SwiGLU, hidden dimension 1664
• Normalization: RMSNorm
• Context length: 2048 tokens, trained natively at this length
• Vocabulary: 16,384 SentencePiece tokens

Training

• Pretraining data: approximately 1.5B tokens of English web and book text
• Sources: FineWeb-Edu, BookCorpus, OpenWebText, PG-19 and WikiHow
• Training hardware: a single NVIDIA Jetson with 8GB of unified memory

Usage

ollama run azerdsq/g1-nano-base "The city of Paris is"

This is a base model: it completes text rather than answering questions.

Weights are also available on Hugging Face.

Limitations

• 60M parameters impose a hard limit on factual knowledge; expect fluent but frequently incorrect completions on knowledge-intensive prompts.
• Maximum context length is 2048 tokens, which remains short compared with modern language models.
• English-only training data.
• Single-sequence generation only; padded batched inference is not supported by the custom model code.
• No instruction tuning and no chat format.

License

Apache 2.0. This release contains model weights and the code required to load them; it does not include the training data or private training infrastructure.