Ultrafast for GPT-6.2 Sol just landed in Codex on the $200/mo Pro plan. Could it give local inference an “ultrafast mode” of its own? I only quantize when a model won’t fit in memory, so the experiment used Qwen3.8-27B’s original BF16 weights with lossless packing. On an M2 Ultra, it hit 51.63 tok/s in Chinese chat and 44.80 tok/s in code completion. Plenty fast for my day-to-day use.