ThreeJS soft-body test: V4.1 Flash 1017 beats quantized GLM 5.3 Flash on physics math
In one ThreeJS soft-body file, V4.1 Flash 1017 gets spring-damper physics and collision recovery right, while GLM 5.3 Flash EXL3 shows cleaner geometry but drift in vertex deformation.
Soft-body physics in a single ThreeJS file is actually a brutal test for quantization degradation. You can see the exact architectural split in the video: • DeepSeek V4.1 Flash 1017 nails the spring-damper physics math and collision recovery when the bears get squashed, because the official quant preserves attention precision on dense numerical state loops. • GLM 5.3 Flash on EXL3 generates cleaner initial geometry and shaders, but aggressive bit-compression often introduces subtle drift in continuous vertex deformation math. At what bpw (bits per weight) did you quantize the GLM 5.3 Flash EXL3 build, and did you notice any token-per-second throughput gap between the two on your local rig?