Today's Qwen 3.8 Flash Next NVFP4 TP=1 DGX Spark update includes native 4096×4096 image processing
Multi-image history, xhigh thinking by default, reasoning retention, and a 4 GiB shared-memory image cache. Three hours of sustained agentic OpenCode work, including visual tasks
Ready! Is there a brave Apple Silicon user with a 64GB machine willing to test DwarfStar with Qwen 3.8 Flash Next?
In the video it runs at 65 t/s on M3 Ultra. 🚀 I uploaded Q2 weights here: PR branch is here:
Why you need a propper RIG?
At the same time Qwen 3.8 Flash Next is cooking voxel viking village on 2x3090, Comfyui is having fun with MiniMax H3 elves on another 3090 :) 130GB LLM simultaneously with next level video generation. Sick :) ps. one 3090 free, but we can always
Qwen 3.8 Flash Next launched 19 agents over the course of 11 hours, and everything is running smoothly
Most of the agents have completed their tasks without errors so far. For me, it’s the best local model for long tasks to date; for short tasks, the 27B model is better because
You have reached the end of the archive
All of qwen38