GLM-5.3 Flash is probably the best agentic coding model you can run on 2× NVIDIAAI DGX Sparks right now (TP-2).
Both Qwen 3.8 Flash Next and Deepseek V4 Flash Vision are faster (and also excellent quality for most workloads), but our opinion is that GLM is just at a different
Qwen 3.8 Max just showed what long-running AI actually looks like.
This isn’t “ask a question and get an answer.” In one competition, the model: → Worked for 24 hours with no human help. → Read the rules itself. → Wrote and tested its own code. → Made 45 submissions.
The trees can be fixed with re-prompting but my objective was to try and get it done in one shot.
Here is a Voxel Pagoda made with Qwen 3.8 Flash Next using laurent_zw 's quant with a Vulkan backend on Strix Halo avg 41 tok/s. Speed burst up to 47 tok/s using ROCmFPX build.
Testing Qwen 3.8 Flash running on 256gb m3 ultra.
Told it to make Halo 3 slayer with bots, make the map Guardian. Gave it reference images for the battle rifle and map. This took 2 hours and sat at about 120gb memory usage
You have reached the end of the archive
All of qwen38