2 x 4090s running Qwen 3.8 27B at a total of over 300 tok/s.
8 concurrent sessions with vLLM, Qwen3.8-27B-INT8-W8A16-MTP, 128k context. Over 1 million tokens per hour for my AI agents to use. Without paying for subs or APIs. Without giving my data away to AI companies.
If you're testing Qwen 3.8 Max, don't waste it on tiny prompts.
Use it where the long context actually matters. Try these: 1. Feed it your full company knowledge base. 2. Give it several SOPs at once. 3. Add your past customer conversations. 4. Include common objections and
GPT turned six product photos into an actual brand.
I gave GPT, Claude, Qwen, and Grok the same six images and the exact same ecommerce brief. GPT-5.6 Sol: ~11m · ~19K tokens · ~$0.56 Grok 4.6 xhigh: 16:34 · 114K · $0.68 Qwen 3.8 Max: ~25m · ~45K · ~$0.60 Claude Opus 5 xhigh:
Calling it a win because computer is stable with 3080 @ 200W, and Qwen 3.8 2-bit made it work
You have reached the end of the archive
All of qwen38