i run Qwen3.5-122B-A10B-Q4_K_L (Bartowski) on strix halo, 220k context slaps. speed falls off a bit with context for sure but i dont care, reasoning quality over everything. 8-20 tok/sec depending on context heaviness

my meta thesis and also that of @tzhen is that this is where the industry is failing right now, specifically we do not have the right r&d loop yet (ideas, (optional/parallel) formalization, data, validation, prototyping, principled comparison) academic artifacts are too disconnected from reality because reality is hard to model. industry products are not principled because they can only design for the burning needs of existing customers, and the dumber less principled ideas win that what we try to create at Flashbots is a middle ground... a soup of both, knowing what we have observed in many systems, and suggesting fixes that address the core issue. for example this piece writings.flashbots.net/mev-a… where we used half a decade of experience dealing with all forms of MEV spam to propose a market based solution that was based on auction R&D we have been doing for years the closed loop approach of a paper probably will not capture the needed complexity for Solana. they will end up with a forest of papers, and products, and protocols, all feeding each other and converging, naturally, because nothing else works. and it will end up looking a lot like what we have been calling SUAVE, because ultimately, we are all building the same thing