M3 Ultra 256GB rapid-mlx: Qwen3.5-27B 4bit cold TTFT

rapid-mlx (MLX) eval scorecard, Apple M3 Ultra 256GB, auto-generated 2026-03-05. enable_thinking false, cache cleared between suites. The scorecard prints two decode rates per model, short and long, without stating the token counts behind them, so only the long figure is recorded and the unit says so. TTFT is the cold figure from the per-model block, converted from the printed seconds where the table used seconds. Single machine, one model resident at a time; per-request, not an aggregate across concurrent users. The exact prompt length for the cold TTFT is not printed. This record is the COLD TTFT. Warm TTFT is printed in the same block and is not filed.

Cold TTFT from the per-model block. 1.3s and 1.1s rows were printed in seconds and are stored as 1300 and 1100 ms.