Update on the free community Qwen 3.8-27B endpoint (was getting a bit too slow)
2x H200 replicas (+1 replica) now use speculative decoding (70 → 126 tok/s, measured) default thinking is now medium when not set (xhigh burns an enormous amount of tokens)
Been using Qwen 3.8 27B (Q4) locally on 64GB of VRAM.
Here is the verdict: SLOW 18 tps with ZERO system prompt to process and that degrades significantly with a harness system prompt and as the context window grows. RIP if you have to compact. I had it implement this PRD
Absolutely cooked with their models.
Since I'm pretty lame I've asked to install the Qwen 3.8-27B for me on my 1080... It got installed and the speed was ~1.8 tok./s. I've then decided to ask Daybreak Blue High to find ways to improve the speed to something more
How to run Qwen 3.8 27b (Opus 4.6) At Home 🏆
Pushing 30+ tk/s on Mac M2 Max // 97 tk/s on 4090 TL;DR: Using Codex + GPT‑5.6 SOL at xHigh reasoning, I stress-tested the new Qwen3.8‑27B across my local hardware and nearly doubled its speed through native speculative decoding.
You have reached the end of the archive
All of qwen38