Qwen 3.8 27b FP8
Xhigh vs medium through Deepseek Harness. Not bad, not bad at all. There's redeeming qualities about both honestly. With a but more cooking in the oven I think either one could be made to look even better.
Indeed. Let’s develop a bit more the top AI developments of this year
1. Open Claw brought to us the free agent at home concept, which Hermes improved upon (more stable). Perfect to combine with free models. 2. GrokBot gave us the best non-free agent which becomes a personal
XERJ + Qwen 3.8... give a cheaper model better context and it starts working smarter and faster
Opus 5.5 and Qwen 3.8 get the same task of building an interactive website, with XERJ helping Qwen find the right context and make fewer guesses as it works. Qwen + XERJ finishes in
100+ tok/sec!!!! Dual Intel B70 32 GB & Qwen 3.8 27B.
64 GB of VRAM, real work, really 109 tok/sec decode and 1,700 prefill. this is real, local, and affordable.
You have reached the end of the archive
All of qwen38