SURPRISE model for the low VRAM folks! Qwen3.5-9B-DeepSeek-V4-Flash is live!
Compared to the base 9B, this DeepSeek-V4 distill wins by a country mile in two specific places:
Reasoning: base overthinks and hits the 8K thinking cap on 3 of 5 prompts; distill clears all 5 cleanly. 2.2× faster time, 2.6× less reasoning length.
Creative front-end design: On creative prompts, the base ships flatter visuals with overlay/animation bugs; distill produces output that punches well above a 9B. See it all for yourself in my full write-up and interactive space! Link in comments! Base and distill raw outputs are presented so you can draw your own conclusions!
Tool calling: 5/6 PASS on both, the fine-tune didn't break tool calling!
Throughput: 143 tok/s flat on both with a 5090, but you could run this model on pretty much anything!
This was a test on our new training pipeline with the Asus GX10 unit, and having confirmed success with this fine-tune, we've already launched the Qwopus 3.6 27B training, which will be completed soon!
It astounds me what we can do with an incredibly clean dataset, a decent base model, and a GX10. You'd think improvement over highly funded lab offerings would not be possible, but here we are!
This is the first model fully completed in the Wyoming lab!
Yeehaw!
https://t.co/QGB88vPZIo
OpenClaw 2026.4.27 🦞
🧠 DeepInfra provider
📎 better file attachments
🛡️ operator-managed proxy routing
🧭 stricter model selection + local model fixes
🔧 gateway, channel, and session reliability
Ships more than it brags.
https://t.co/G3XvH9OQWm