I’ve been running it for about two weeks on a single B70, and it has become my daily driver
My setup is pretty modest with
Qwen3.8-Flash-Next (Swift 1.5 IQ3_S). The experts mostly live in system RAM and the B70 holds the hot ones, I compared this to Qwen 3.8 27B and ended up sticking with this
The main driver was the performance difference between Strata and XeStrata
I had claude create some arbirtrary workloads for a quick comparison and got the following results
Absolute numbers for both models through llama-swap, 192k context, INT8 KV, MTP spec 4)
Strata vs XeStrata (same GGUF, context, MTP, prefill and reserve, 3 seeds, thinking on)
Decode tok/s, upstream Strata v0.1.45 → XeStrata:
┌──────────┬────────┬──────────┐
│ Task │ Strata │ XeStrata │
├──────────┼────────┼──────────┤
│ bugfix │ 37.7 │ 51.4 │
├──────────┼────────┼──────────┤
│ code │ 32.7 │ 49.2 │
├──────────┼────────┼──────────┤
│ invoice │ 31.7 │ 66.2 │
├──────────┼────────┼──────────┤
│ incident │ 28.7 │ 52.9 │
└──────────┴────────┴──────────┘
Speed was there but should be noted that these are pretty much one shot runs, not an average over a large number of prompts. Answers on the checkable tasks were equally correct. The text was not token-identical
Conversation cache
This is the feature I like most for agentic use. The engine parks finished conversations’ KV state in RAM I give it 25 GiB, and retain up 16 conversations. When a conversation comes back, it resumes instead of re-prefilling:
For agents and coding tools that re-send a long prefix every turn, this matters more than raw decode speed. Things I hit