Skip to content

Model Recommendations

Use this page when you want a conservative starting point for MLX-Gen model selection on Apple Silicon Macs with 18 GB, 24 GB, 32 GB, 64 GB, or 128+ GB unified memory.

These recommendations are based on published MLX-Gen validation profiles, not on package size alone. When a benchmark includes full-process Darwin physical peak memory, that is the number used for tiering. Otherwise this page falls back to the closest published peak RSS figure for the same route. Larger canvases, more denoise steps, longer videos, and heavier editing modes can require more memory than the example profile shown here.

Use the linked topic pages for the full command surface and benchmark details:

Memory tier Text-to-image Image-to-image and edit Video generation Upscaling and video restore
18 GB Start with prism-ml/bonsai-image-ternary-4B-mlx-2bit, AbstractFramework/flux.2-klein-4b-8bit, or AbstractFramework/ernie-image-turbo-8bit. Use ernie-image-turbo-8bit for latent restyle and flux.2-klein-4b-8bit for lighter edit/reference work. Avoid Qwen edit and control routes here. No production recommendation. Bernini-R 1.3B fits the tested low-RAM memory shapes but remains experimental despite accepted official-example proof at full profile. seedvr2-3b and seedvr2-7b image upscaling packages fit comfortably. Treat published SeedVR2 video restore as a higher-memory workflow.
24 GB Add AbstractFramework/fibo-8bit, AbstractFramework/z-image-turbo-8bit, and AbstractFramework/qwen-image-2512-8bit text-only runs. z-image-turbo-8bit latent and native inpaint become practical. flux.2-klein-4b-8bit remains the lighter edit default. No conservative Wan A14B recommendation since the 2026-06-12 q8 runtime-precision fix (A14B now needs BF16-class memory, ~33 GiB physical peak). SeedVR2 image upscaling remains easy. The published five-second video-restore profile still needs more headroom than this tier.
32 GB The 24 GB image recommendations remain good with more headroom. qwen-image-2512-8bit text-only runs are comfortable. For edit-heavy work, z-image-turbo-8bit and flux.2-klein-4b-8bit are still the safer defaults than Qwen Edit. Wan A14B runs at BF16-class memory since 2026-06-12 (~33 GiB physical peak at the small low-RAM profile), above this tier's conservative envelope. seedvr2-3b-8bit published 29/8 five-second video restore is in range; seedvr2-7b video restore is still above this tier in the published profile.
64 GB All image families above are practical. Add AbstractFramework/qwen-image-edit-2511-8bit for strong layout-preserving edit/reference work, plus AbstractFramework/qwen-image-8bit control and control-inpaint. Wan A14B (q8 storage or BF16; both run at BF16-class memory, ~33 GiB physical peak in the small low-RAM profile) is the first conservative tier for A14B video. seedvr2-7b image upscaling is easy. For published video-restore proofs, keep using seedvr2-3b unless you have more than 64 GB free.
128+ GB Everything above. Everything above, plus room for larger or more exploratory runs. Add native AbstractFramework/wan2.2-ti2v-5b-diffusers-8bit high-resolution video profiles such as 1280x704, 17 frames, 20 steps. Add published seedvr2-7b five-second 29/8 video restore and larger Wan experimentation with real headroom.

Benchmark Basis

Route Model Published profile Peak memory used for recommendations
T2I prism-ml/bonsai-image-ternary-4B-mlx-2bit 512px, 4 steps 3.57 GiB peak RSS
T2I AbstractFramework/flux.2-klein-4b-8bit 512px, 4 steps 9.23 GiB peak RSS
T2I / latent I2I AbstractFramework/ernie-image-turbo-8bit 512px, 8 steps 12.9 GiB peak RSS
T2I AbstractFramework/fibo-4bit 512px, 8 steps 11.39 GB max RSS
T2I AbstractFramework/fibo-8bit 512px, 8 steps 15.89 GB max RSS
T2I AbstractFramework/qwen-image-2512-8bit 768x336, 4 steps 10.73 GiB peak RSS
Latent I2I AbstractFramework/z-image-turbo-8bit 768x432, 9 steps 11.49 GiB peak RSS
Native inpaint AbstractFramework/z-image-turbo-8bit 768x432, 9 steps 10.57 GiB peak RSS (improved by the 0.23.0 denoise dtype fix; pre-fix value was 18.11 GiB)
Edit/reference I2I AbstractFramework/qwen-image-edit-2511-8bit 768x432, 4 steps 30.91 GiB peak RSS
Control-inpaint AbstractFramework/qwen-image-8bit 768x432, 4 steps 34.94 GiB peak RSS
Image upscale AbstractFramework/seedvr2-3b-8bit 5x from 133x113 source 4.73 GiB max RSS
Image upscale AbstractFramework/seedvr2-7b-8bit 5x from 133x113 source 8.90 GiB max RSS
Video restore ByteDance-Seed/SeedVR2-3B five-second 149 frame 29/8 published profile 27.40 GB max RSS
Video restore ByteDance-Seed/SeedVR2-7B five-second 149 frame 29/8 published profile 66.18 GB max RSS
T2V AbstractFramework/wan2.2-t2v-a14b-diffusers-8bit 384x224, 33 frames, 12 steps, low-RAM 33.0 GiB physical peak (BF16-class since the 2026-06-12 q8 runtime fix; pre-fix value was 20.7 GiB)
I2V AbstractFramework/wan2.2-i2v-a14b-diffusers-8bit 384x384, 33 frames, 12 steps, low-RAM 33.7 GiB physical peak (BF16-class since the 2026-06-12 q8 runtime fix; pre-fix value was 21.5 GiB)
T2V / I2V AbstractFramework/wan2.2-ti2v-5b-diffusers-8bit 1280x704, 17 frames, 20 steps 103.7 GiB physical peak
R2V ByteDance/Bernini-R-1.3B-Diffusers factored BF16 320x192, 17 frames, 20 steps, 8 references, low-RAM 9.12 GB sampled physical peak
RV2V ByteDance/Bernini-R-1.3B-Diffusers factored BF16 176x320, 17 frames, 20 steps, source + reference, low-RAM 9.22 GB sampled physical peak
V2V control ByteDance/Bernini-R-1.3B-Diffusers factored BF16 176x320, 17 frames, 20 steps, source only, low-RAM 9.45 GB sampled physical peak

Practical Reading

  • 18 GB is enough for compact image families and SeedVR2 image upscaling. Bernini BF16 low-RAM fits the tested shapes (9.45 GB maximum bounded-matrix peak; 8.17 GB for the 33-frame, eight-reference 848-condition structural probe), but it is not recommended for production: its visual release gate fails. The official 81-frame profile remains unmeasured, and a fresh selective download separately needs 18.36 GiB free disk including headroom.
  • 64 GB is the first conservative tier for Wan A14B video: since the 2026-06-12 q8 runtime-precision fix, A14B runs at BF16-class memory (~33 GiB physical peak even at the small low-RAM benchmark profile).
  • 32 GB is the first tier where the published five-second SeedVR2 3B video-restore profile fits the published memory envelope.
  • 64 GB is the first tier where Qwen edit and Qwen control routes become practical without running at the ceiling.
  • 128+ GB is where the current native TI2V-5B high-resolution benchmark belongs.

Not Yet Published As Memory Recommendations

MLX-Gen has additional validated quality surfaces that are not promoted here because the current docs do not yet include a strong enough route-specific memory benchmark for a conservative public recommendation. That includes FLUX.2 Klein 9B, Qwen q4 edit packages, and broader Z-Image q4/q8 route combinations outside the measured rows above.