@accelerate27 Sorry for the late reply, I somehow missed the notification.
Thanks for the comparison, Qwen4-Max does look more solid in 3D.
(The recording is a bit rough though. :D)
I tried this prompt with Qwen4 Max Preview.
Not 1 turn.
After the first turn it generated very patchy output (stray blocks everywhere)
Then another turn because it put trees in the front side obstructing what the camera can see + told it to scale up because even the people look like 2-3 blocks tall. But it did not work. So I stopped here.
Total 3 turns.
Method used: I used Qwen4 Max Preview Xhigh as orchestrator to orchestrate Qwen4 Max Preview Xhigh subagents
GLM 5.3 Flash on two CMP 170HX GPUs in vLLM, now ~2× faster at the same size and accuracy 😎
I've been having some fun with this one.
Updated benchmarks:
• Decode: 111–176 tok/s single stream, up to 428 tok/s at 4 concurrent
• Prefill: 4,100–4,600 tok/s
• 225K context, image + video
The trick: Q4NX, a new EXL3 codebook built specifically for these cards 🧵