Pyrite: Byte-Neutral Group-Size Trading for 2-Bit Quantization of 27B–33B Language Models on 16 GB Apple Silicon

Running 27B–33B parameter language models on a 16 GB unified-memory laptop requires roughly 2-bit weights, at which point quality degrades sharply. We describe Pyrite, a simple, kernel-free recipe for standard affine group quantization (as implemented in Apple MLX) that redistributes bits without changing the total tensor size by a single byte. The recipe widens the quantization group of the SwiGLU gate_proj/up_proj matrices from 64 to 128 and reinvests exactly the freed bytes to upgrade a sensitivity-ranked subset of projections from 2-bit to 3-bit. Because the cost of each move is an integer multiple of a common unit, any allocation satisfying 4*n_down + 5*n_gate/up = 2L (for L layers) yields a byte-identical model. On a sparse Mixture-of-Experts model (llm-jp-4-32b-a3b-base) the best configuration reduces held-out Japanese KL divergence to the BF16 teacher by 11.0% (English: 12.9%), and on a dense hybrid model (Qwen3.8-27B) by 22.3% (English) and 22.4% (Japanese), relative to uniform 2-bit quantization of identical size. In both architectures, allocations that mix down_proj upgrades with a small number of late-layer gate/up upgrades outperform allocations that upgrade down_proj only. All profiling and evaluation run on a single 16 GB M2 Pro by streaming one layer at a time. The resulting models load with unmodified mlx-lm. Version 2.0 adds empirical validation on the international flagship coding MoE: deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct (15.7B total, 2.36B active), achieving 4.88 GiB under exact zero-byte overhead with improved multi-domain validation perplexity and token accuracy.