got deepseek V4.1 flash running locally on a 16GB m1 mac mini
original FP4/FP8 weights, ssd streaming + custom mlx runner
108s ttft and about 23s/token (not to be confused with tok/s)
goddamn, CursorBench 3.2:
fable 5.1 medium: 68.0% — $3.53/task
gpt-5.6 sol max: 67.2% — $5.69/task
fable is now beating sol while costing way less per task
i said this efficiency jump was coming, just didn't expect it to get this cheap this fast
img2threejs - turn one object photo into a code-only procedural Three.js model
Open-source toolkit that rebuilds the object in a single reference image as procedural Three.js code (no mesh downloads), quality-gated by a render-vs-reference loop - strong for hard-surface objects