I didn't collect hard data, but speed was about third of Codex with 2 4090 at full wattage, 128k context with kv. My main problem was the quality. Qwen, Gemma outputs were buggy and clearly out of sync with what the repo already have. I only had 2 days with the hardware to test, so not sure if I could have done better.
None of the local models get anywhere near the productivity I get with Claude or Codex. I even borrowed a stack of GPUs to run some tests last month and it was still too slow and too dumb. I feel like manually coding with a good IDE is more productive than any local model.
Did you reset the index and search relay lists? If it is taking that long it's because the app cannot find the outbox relays of your follows, which comes from index and search. Just reset those and it should be good to go.