Found a sweet spot between quality and speed for Qwen 3.8 Flash on 2x3090s @ 3.5bpw, running some tests
Backend is exllamav3, running pretty sweet @ around 105 tk/s average (92 tk/s prose/reasoning, 105 tk/s coding and 120 tk/s file editing) and 1635 tk/s prefill
DeepSeek V4 Flash, Qwen 3.8 27B, GLM 5.2, Nemotron 3 Ultra, Laguna S 2.1, and Jev for agent routing.
All FREE: OpenRouter: DeepSeek paths, Qwen 3.8 27B, GLM 5.2, Ling 3.0 Flash VL, Laguna S 2.1, Nemotron 3.5 Lightning / 3 Ultra: OpenCode Zen:
Qwen 3.8 Max Full COURSE 1 HOUR (Build & Automate Anything)
Want the SOP? DM me. 💬
You have reached the end of the archive
All of qwen38