PCEI is a GitHub project that presents Physically Constrained Emergent Intelligence as an alternative to scaling monolithic Transformer models. Its claimed design combines a 24-layer Mamba selective state-space backbone, 1.58-bit ternary weights constrained to -1, 0, and +1, a dual-timescale ambient field, and a two-stage wakeful-and-dreaming training lifecycle. The repository says the 750M-parameter system can run on a single 24GB consumer GPU or Apple Silicon, with linear sequence complexity and a constant recurrent state during inference. It reports a ternary weight distribution of 40.08% -1, 19.82% zero, and 40.09% +1, and claims integer-addition computation can reduce matrix-operation energy use by up to 10 times. Its latent diffusion dreaming procedure is reported to reduce held-out perplexity from 57.79 to 38.72 after 1,000 cycles without external training text. Other documented experiments report 1,475 tokens per second at sequence length 2,048 with 54% less VRAM, fluency training on 334 million tokens, and an evaluation-loss reduction from 6.40 to 4.72 after adding global-workspace broadcasting. The project includes 11 research papers, code for multimodal processing and local inference, and scripts for consolidation and diagnostics. It requires Python 3.10 or newer and PyTorch 2.2 or newer, and is distributed under an MIT Non-Commercial license.
