On Strix Halo, have a look at the work antirez is doing with dwarfstar (antirez/ds4) for deepseek v4 flash at reasonable speeds and quality.
Laguna S 2.1 runs at 15-28 tokens per second, depending on context and...something about how long it's been running, which is very comfortable for chatting, but still not usable for interactive agentic coding. Their `pool` agent just times out when I try to use it with the Strix Halo-hosted instance of the model.