[baq]

the problem is you're still a year behind with this approach and it isn't at all clear locally hosted models can keep the gap. need more turboquant-like algorithmic boosts for this to happen.