Benchmarking five local LLM inference engines on an M1 Max 64GB with Qwen3.5-9B. Comparing GGUF and MLX decode throughput, speculative decoding gains (MTP vs DFlash), and concurrent request scaling.
Nova Anime AM, CottonAnima, AMIXA2.9B, and JANIMA 2.9B tested on an M1 Max 64GB in ComfyUI with an Anima-Base character LoRA remapped to 40 blocks. All four kept the hairstyles, but Kei's red ribbon tie turned navy in portraits. Also compares default style, background density, i2i, and speed.
I ran the latest 2B versions (as of October 21) of Hikari, AnimaIka, Hotaru Blend, Cat Tower, CapAnima, RDBT, and AnimaYume in ComfyUI on an M1 Max 64GB, using the same prompts, seed, and character LoRA as Anima-Base v1.0. All seven got the hairstyles right, but only RDBT kept Kei's ribbon tie red in both full-body and bust shots, and RDBT turned two characters into four. Default style, background density, i2i, and generation speed are compared too.
I compared official Anima Turbo v1.1 and Aesthetic v1.1 against Base v1.0 on an M1 Max 64GB with ComfyUI. Base-trained character LoRAs worked across all 4 seeds when specifying hair and eyes, but Aesthetic v1.1 spawned extra characters in two-character prompts and Turbo v1.0 missed ribbon colors.
A local evaluation of MOSS-Image Base and SFT on an M1 Max Mac using MPS, examining prompt behavior for character portraits, text rendering, JSON layout placement, and comparisons with earlier Qwen and Anima generations.
We evaluated MiniCPM-V 4.7 35B-A3B on a 16GB M4 Mac mini across coding, reasoning, and table OCR benchmarks, finding 100% transcription accuracy alongside arithmetic hallucination patterns.
Tested on M1 Max: Diffusers stretches an 832×1216 reference to 832×1248, making figures 2.7% narrower. Setting output_resolution=1006 cut whole-image drift to about 0.1px.
Tested on M1 Max 64GB: the training-free TaylorSeer cache in Diffusers made Qwen-Image 2.1 about 2.7x faster. It crashes with the default KV cache; a small patch fixes it.
Tested Qwen-Image 2.1 open weights on M1 Max 64GB: overhead shots work locally, ~10 min per 832×1216 image (2x+ Anima), and i2i kept the face. Same prompts as my Anima/WAI tests.
Tested on M1 Max 64GB: AtomicChat's M64 GGUF keeps the 51B N-gram table in its own shard, so a llama.cpp PR #27742 build leaves it on SSD and runs the 125B MoE at 17.6 tok/s.
Tested on M1 Max 64GB with MLX: LLaMA Pro-style block expansion on Qwen3-0.6B-Base vs LoRA vs full FT, trained on 722 blog posts. Held-out PPL drops 14.7 to 10.5, but Wikipedia-ja rises 13.5 to 17.9, more forgetting than full FT. A 20% Wikipedia mix nearly removes it, and renumbering 28-layer LoRA keys reproduces the Anima-2.9B result in numbers.