Results.
What changes when precision follows entropy: quality, where tokens land, what rounding costs, and what it does to speed.
In low gear, weights take less memory than fp16.
On predictable passages, generation runs up to faster.
Against full precision, quality on standard benchmarks drops by .
Quality
Perplexity on WikiText-103 (lower is better). Static 4-bit falls apart; routing stays close to fp16 because the tokens that matter run at full precision.
Where tokens land
Share of generated tokens in each gear, by kind of prompt. Code spends the most time in the cheap gears; open-ended explanation spends the least.
What rounding costs
Quantization error measured on the model's real weights.
Speed and shifts
Decode throughput by gear and the cost of switching gears.
Method
Every number on this page comes from paper/results/measurements.json in the repository and is regenerated by paper/run_paper_benchmarks.py, run_memory_bench.py and run_perplexity.py. Runs use fixed seeds and report the median of repeated runs. Perplexity uses 8,192 tokens of the WikiText-103 test split in 512-token chunks.