ModelFit Lab run record: LTX-2.3 and MiniMax H3 video on a Mac mini M5 Pro 48 GB Run date: 2026-11-04 (Asia/Bangkok). Machine - Mac mini, Apple M5 Pro, model Mac17,16, 48 GB unified memory (hw.memsize 51539607552), macOS 27.2 - Metal recommendedMaxWorkingSetSize 40200.90 MB (reported by sd-cli --list-devices), iogpu.wired_limit_mb 0 - Weights and outputs on an external USB4 SSD. Each run went alone through the shared mq gpu queue (one GPU job at a time). - One run per setting. Time and memory come from the runner logs and /usr/bin/time -l on this Mac. Memory columns - peak footprint: "peak memory footprint" from /usr/bin/time -l, in bytes, with GB = bytes / 1e9 - max RSS: "maximum resident set size" from /usr/bin/time -l - swap: vm.swapusage "used", sampled every 2 s during the run - With --low-ram (LTX) and --params-backend disk (stable-diffusion.cpp) weights are read from memory-mapped or re-read model files, so those file pages are not part of the peak footprint. LTX-2.3 runs - Runner: ltx-2-mlx (github.com/dgrauet/ltx-2-mlx, commit 895201df379b079c2154738bcf92005b7a371923, ltx-pipelines-mlx 0.18.0, ltx-core-mlx 0.18.0), mlx 0.32.2, mlx-metal 0.32.2 - Text encoder: mlx-community/gemma-3-12b-it-4bit (default of ltx-2-mlx) - Command: ltx-2-mlx generate --model --distilled -H -W -f 121 --frame-rate 24 --seed 42 --prompt -o [--low-ram] - Two-stage distilled pipeline, 121 frames at 24 fps (about 5 s), video with audio - Time: the "Time:" line printed by ltx-2-mlx (generation, after the process starts) | id | pack | flags | resolution | frames | fps | time (s) | wall (s) | peak footprint | max RSS | swap | | ltx-2-3-q4-704x448 | dgrauet/ltx-2.3-mlx-q4 (int4) | none | 704x448 | 121 | 24 | 74.2 | 74.48 | 17524845464 bytes (17.52 GB) | 13049577472 bytes | 1230 MB before and after, no growth | | ltx-2-3-q4-1280x704 | dgrauet/ltx-2.3-mlx-q4 (int4) | none | 1280x704 | 121 | 24 | 232.7 | 233.02 | 19316483384 bytes (19.32 GB) | 13599932416 bytes | 1230 MB before and after, no growth | | ltx-2-3-bf16-lowram-704x448 | dgrauet/ltx-2.3-mlx (bf16) | --low-ram | 704x448 | 121 | 24 | 92.6 | 93.04 | 15710841896 bytes (15.71 GB) | 14251458560 bytes | 1869 MB before, 1813 MB after, no growth | | ltx-2-3-bf16-lowram-1280x704 | dgrauet/ltx-2.3-mlx (bf16) | --low-ram | 1280x704 | 121 | 24 | 245.1 | 245.38 | 15702535208 bytes (15.70 GB) | 14251622400 bytes | 2055 MB before, 2023 MB after, no growth | | ltx-2-3-bf16-704x448 | dgrauet/ltx-2.3-mlx (bf16) | none | 704x448 | 121 | 24 | 98.8 | 99.08 | 41201004336 bytes (41.20 GB) | 19355484160 bytes | swaps on 48 GB: 1813 MB before, 4611 MB peak, 2450 MB after | Pack revisions: dgrauet/ltx-2.3-mlx at e70722e4c21ad04b47ef7b6cfb433f9cbf42d92e (transformer-distilled-1.1.safetensors), dgrauet/ltx-2.3-mlx-q4 (transformer-distilled-1.1.safetensors, README identical to main on 2026-11-04). Shared files (connector, VAEs, vocoder, spatial upscaler x2) are byte-identical across the packs (same LFS sha256 on Hugging Face). MiniMax H3 runs - Runner: stable-diffusion.cpp sd-cli, release master-1047-984eb06 (macOS arm64 binary, Metal), commit 984eb06 - Weights: unsloth/MiniMax-H3-GGUF at babc040eff0084465460ead58fdb8ec44296f8dd: minimax_h3_fl2va_pruned-Q4_K.gguf, qwen3vl_32b_minimax_h3-Q4_K_M.gguf, vae/minimax_h3_video_vae_fp16.safetensors, vae/minimax_h3_audio_vae_fp32.safetensors (the two VAEs have the same LFS sha256 as Comfy-Org/MiniMax-H3) - Command: sd-cli -M vid_gen --diffusion-model minimax_h3_fl2va_pruned-Q4_K.gguf --llm qwen3vl_32b_minimax_h3-Q4_K_M.gguf --vae minimax_h3_video_vae_fp16.safetensors --audio-vae minimax_h3_audio_vae_fp32.safetensors --cfg-scale 1.0 -W 352 -H 608 --fps 24 --video-frames 124 --seed 77001 -p - Text to video with audio, Euler, 20 steps (sd.cpp default), no turbo LoRA, cfg 1.0 - Time: "generate_video completed in" printed by sd-cli (text encoding, sampling, audio and video decode) | id | memory flags | resolution | frames | fps | steps | outcome | time (s) | wall (s) | peak footprint | max RSS | swap | | minimax-h3-q4-k-352x608-disk | --params-backend disk --rng cpu | 352x608 | 124 | 24 | 20 | success | 1029.63 | 1031.22 | 19592346240 bytes (19.59 GB) | 19593576448 bytes | 2450 MB before and peak, 2055 MB after, no growth | | minimax-h3-q4-k-352x608-offload | --offload-to-cpu --rng cpu | 352x608 | 124 | 24 | 20 | failed | none | 51.86 | 40253387912 bytes (40.25 GB) | 20068024320 bytes | 1214 MB before, 13978 MB at stop | MiniMax H3 success stages: text encoding 26.41 s, sampling 955.43 s, audio decode 3.16 s, video decode 44.55 s. Output 352x608, 124 frames, 24 fps, stereo 32 kHz audio. MiniMax H3 failure: with --offload-to-cpu the swap rose from 1214 MB to 13978 MB while the diffusion model loaded after text encoding, and our own swap guard (stop above 8 GB of new swap) ended the run before the first sampling step.