2026-10-05 · America/Los_Angeles · 社区动态 · #7

Qwen3.8-Flash-Next (125B) on a single Strix Halo mini PC: 44-59 tok/s with speculative decoding, ~1,400 tok/s prefill, engine is open

Hey all. We've spent the last weeks getting Qwen3.8-Flash-Next (125B MoE, 6B active) to run properly on one AMD Strix Halo box (Ryzen AI Max+ 395, 128 GB). Tonight we're releasing both the 95 GB EXL3 weights and a new version of Kyojin, our inference engine (built on ExLlamaV3, open). This is a first version, same as our GLM-5.3-Flash and MiMo-V2.6-Flash builds. We'd rather ship it and improve it in the open: speed and quality updates are coming for all three. Numbers, all from a fresh clone and build on the mini PC: Decode: 44 to 59 tok/s with speculative decoding depending on the task…

热度 49.2 / 100；排名与评分保留该期记录。