Home AI Bench lab noteshttps://homeaibench.com/notes/What we tried on home GPUs, including the runs that failed and how we fixed them.enSun, 25 Oct 2026 09:00:00 +0900 MOSS-Image on a 12 GB card: Base, SFT and an anime fine-tune, side by sidehttps://homeaibench.com/notes/2026-10-25-moss-image-anime-comparison/https://homeaibench.com/notes/2026-10-25-moss-image-anime-comparison/Sun, 25 Oct 2026 09:00:00 +0900MOSS-Image is a 5-billion-parameter text-to-image model from OpenMOSS, released in October 2026 under the Apache-2.0 license. Within days there were dozens of d Does raw GPU speed predict image generation time? Six cards, three rounds, and a first answer that did not survivehttps://homeaibench.com/notes/2026-10-24-raw-gpu-speed-vs-sdxl/https://homeaibench.com/notes/2026-10-24-raw-gpu-speed-vs-sdxl/Sat, 24 Oct 2026 09:00:00 +0900Every GPU review gives you a few raw numbers: how many trillion operations per second, how many gigabytes per second of memory. We wanted to know whether any of FLUX.3 [klein] on a 12 GB RTX 5070: five tries, one clean pipelinehttps://homeaibench.com/notes/2026-10-24-flux3-klein-12gb/https://homeaibench.com/notes/2026-10-24-flux3-klein-12gb/Sat, 24 Oct 2026 09:00:00 +0900FLUX.3 klein came out in October 2026. It needs a newer ComfyUI than the container image we use, so each rented node updates ComfyUI to v0.42.1 at start-up. FLUX.2 [klein] 4B: about 3 seconds on a 12 GB cardhttps://homeaibench.com/notes/2026-10-24-flux2-klein-4b/https://homeaibench.com/notes/2026-10-24-flux2-klein-4b/Sat, 24 Oct 2026 09:00:00 +0900FLUX.2 klein 4B (Apache-2.0) runs in 4 steps at cfg 1. Edited open-weight models and hidden backdoors: what is known, and what you can actually checkhttps://homeaibench.com/notes/2026-10-24-edited-open-weights-backdoors/https://homeaibench.com/notes/2026-10-24-edited-open-weights-backdoors/Sat, 24 Oct 2026 09:00:00 +0900A post on X on 23 October warned that open-weight models whose weights have been edited, for example "uncensored" or "abliterated" variants, can hide backdoors. Z-Image Turbo did not fit a 16 GB card in our setuphttps://homeaibench.com/notes/2026-10-23-zimage-turbo-16gb/https://homeaibench.com/notes/2026-10-23-zimage-turbo-16gb/Fri, 23 Oct 2026 09:00:00 +0900Z-Image Turbo (Apache-2.0) is a 12.3 GB model plus an 8 GB text encoder. The 0.1-second trap: ComfyUI caches identical jobshttps://homeaibench.com/notes/2026-10-23-same-seed-cache/https://homeaibench.com/notes/2026-10-23-same-seed-cache/Fri, 23 Oct 2026 09:00:00 +0900Our first warm runs reported about 0.1 s per image. That was not a fast GPU.