I still can't get over how close the qwen 3.8 125b landed to the glm 5.3 flash 320b building the same floating tree.

So watch! here is every number from both runs in one place, serve settings included, save it if you're picking a local model for agent builds glm 5.3 flash

Sudo suOct 57

I had two of the best local models build the same floating tree

Glm 5.3 flash at 320b and qwen 3.8 flash next at 125b, and despite the size gap look where both landed glm 5.3 flash: 320b moe with 18b active, nvidia's nvfp4 weights split over 2x dgx spark, served with vllm

Sudo suOct 540

Qwen 3.8 27B + TensorFold built a game for my Hardball 3D challenge in 3h 14m.

Of the three builds, it’s the most playable, even more so than the GPT-6 Luna baseline. Decode · 90th percentile: 42.37 tok/s Prefill · mean: 160.55 tok/s Run details:

Alfredo Artiles 🇨🇺Oct 52

Paris Nocturne - by Qwen 3.8 Flash Next at 3bpw no thinking loop no break one shot prompt 3 hour working...

This complexity, the thinking process almost same as intelligent with Opus 5.5 or Fable 5.. look at the shading sky, the horizon, the baloon, ferrieswheel, the river, the

lackonendesOct 54