Qwen3.8-Flash-Next+strix halo+halogen server = TAM

Im convinced AI is local now. Halogen loads QFN in 30 GB at 4bits. Its token gen is slightly faster than readable. You can follow along and leave it alone based on its thinking traces.

It gets 256k context but using dynamic context pruning, i got it to loop between 64k and 128k to maximize its speed vs prefill.

Its real. If the memory cartel is busted, the moat is busted.