Parameter-efficient small LLM training experiments achieve <16MB/10 min training using recurrent/shared-block transformers, unigram/bigram logit calibration, and INT8-compressed language model architectures with MLX and PyTorch.
pytorch neural-networks tokenization model-compression distributed-training int8-quantization llm-training-from-scratch parameter-efficient-training
-
Updated
May 18, 2026 - Python