Personally been using GPT-OSS-120b locally with reasoning_effort set to `high` and it blows pretty much every other local model out of the water, but takes a lot of time for it to eventually do a proper content reply. But for fire-and-forget jobs like "Create a well-researched report on X from perspective Y" it works really well.