Tested on M1 Max 64GB: hooked Qwen3.6-35B-A3B's MoE router in mlx-lm, pre-warmed the top-20 hot experts, still ~62 tok/s vs plain mmap cache. Plus the Metal OOM on Qwen3.5-122B.
Tested on M1 Max 64GB: SeFi-Image turbo runs on MPS in bf16 at 13–47s/image; fake text and face artifacts only clear up on 5B RL at 50 steps, 18 min/image.
Measured on M1 Max 64GB: weight-only INT8 runs 41% slower than fp16, a hand-written Metal int8 GEMM 15.6x slower, MPSMatrix int8 3x, and MLX 8-bit 9% slower.
Tested on M1 Max 64GB ComfyUI: v0.24.1 fails to load int8_tensorwise, v0.30.1 hits the missing aten::_int_mm MPS kernel, and a dequantize patch runs slower than bf16.
Tested on M1 Max 64GB: AnimaLoraToolkit + Anima-Base crashes on torch 2.6.0 (MPS SDPA bug) but runs 300 steps clean on 2.7.1, at ~21 s/step, roughly 9x slower than an RTX 5090.
Krea 2 Turbo bf16 renders in ~3.5 min on a Mac (M1 Max 64GB, ComfyUI). fp8 is rejected on MPS, Raw NaNs to black at 47 min, and NSFW behavior gets tested. Working settings and VRAM notes for Apple Silicon.
Tested Boogu-Image-0.1 on an M1 Max 64GB ComfyUI: the fp8 build is rejected by MPS, so bf16 is mandatory, and Turbo runs ~70s per 1024px image. Notes on photoreal vs anime, bilingual text, and where NSFW stops.
Tested on M1 Max ComfyUI: newly free JANIMA vs Hexer Minimal Toon Anima V1 vs anima-base, one character LoRA, same seed. Hexer keeps the outfit; JANIMA adds clothes but draws the quietest backgrounds.
AFM 3 splits into 20B on-device sparse (NAND-to-DRAM weight loading) and Cloud Pro on Google Cloud NVIDIA GPU. Three Google contexts, Foundation Models API opening, and what's still unreleased.
Tested LFM2.5-1.2B-JP-202606 on M1 Max 64GB. llama.cpp Q4_K_M: 208 tok/s decode, JSON intact, model name hallucinated (LFM→FDM). Q8_0: 157 tok/s, no hallucination. Tool calls broken via GGUF.
