Qwen 3.5 122B hits 120K context

This is exactly the kind of benchmark that keeps local inference interesting: not a flashy new release, but proof that aggressive quantization and open-weight models keep stretching cheap secondhand hardware farther than expected.

• –The headline result is less about raw model quality than feasibility: 120K context on dual Mi50s is a strong signal for budget-minded local setups.
• –Prompt processing at ~136 t/s is solid for long-context experimentation, even if decode at ~18 t/s still limits interactive use.
• –The post reinforces how much mileage the open Qwen ecosystem, GGUF quantization, and llama.cpp-style tooling are getting out of non-NVIDIA hardware.
• –Because this is a single community benchmark, developers should treat it as a reproducibility lead, not a definitive performance baseline across workloads.