Efficient LLM Architectures for 32GB RAM (Ternary and Sparse Inference)
Native ternary training beats post-training quantization for memory efficiency.
BitBop: latent-free ternary training of small language models
Fits a 325M model in 6GB VRAM where STE and float baselines crash.
ML researchers and engineers working on model quantization
BitNet · Microscaling formats (MX)
Native ternary training beats post-training quantization for memory efficiency.
From-scratch 1B model training for $315 inverts the massive-budget assumption.
4-month VAE research artifact; reconstruction quality matters less than you'd think.
Open weights for 20 robot embodiments when most VLA models stay closed.
Fixed-latency language-rule decisions beat traditional token-by-token LLM agents.
Train a working LLM in 5 minutes on free Colab with a fish personality.