Can the Mac 24 GB run LTX-2.3 (22B)?
Download Q4_K_M (14.3 GB). With the model's working memory it needs about 19.1 GB, leaving 0.5 GB spare on 19.6 GB. Quality: good; some loss in fine detail and text. For better quality, Q5_K_M (16.1 GB) also runs, with about 1.3 GB spilling into system RAM — a little slower, still practical.
01What to download
| Part | File | Folder | Size | |
|---|---|---|---|---|
| Model | ltx-2.3-22b-dev-Q4_K_M.gguf Q4_K_M · unsloth/LTX-2.3-GGUF | models/unet | 14.3 GB | Download → |
| Text encoder | gemma_3_12B_it.safetensors Gemma 3 12B BF16 · Comfy-Org/ltx-2 | models/text_encoders | 24.4 GB | Download → |
| VAE | ltx-2.3-22b-dev_video_vae.safetensors LTX-2.3 video VAE · unsloth/LTX-2.3-GGUF | models/vae | 1.5 GB | Download → |
| Also needed | ltx-2.3-22b-dev_audio_vae.safetensors LTX-2.3 audio VAE · unsloth/LTX-2.3-GGUF | models/vae | 0.4 GB | Download → |
| Also needed | ltx-2.3-22b-dev_embeddings_connectors.safetensors LTX-2.3 text embeddings connectors (projection) · unsloth/LTX-2.3-GGUF | models/text_encoders | 2.3 GB | Download → |
| Total download · keep about the same free on disk | 42.8 GB | |||
Only for some workflows: LTX-2.3 spatial upscaler x2 1.1 (1.0 GB, official two-stage template). On a Mac, FP8 text encoders do not load, so a 16-bit one is listed. Sizes read from Hugging Face (2026-09-25). “Download” links start the file directly; the file name links are the same files the official ComfyUI workflows use.
Shared memory: about 48.8 GB of your 24 GB. On this machine the model, the text encoder, the VAE and the operating system all use the same memory. That is about 24.8 GB more than there is. It still runs: while the model works, the system compresses or swaps the idle text encoder, so loading and changing the prompt get slower. A smaller file or text encoder avoids that. calculated
02Every LTX-2.3 file on 19.6 GB
| File | Size | Needed | On this card | Quality | Download |
|---|---|---|---|---|---|
| 16-bit GGUF · unsloth | 42.0 GB | 46.8 GB | 27.2 GB too big | the original weights | Hugging Face → |
| Q8_0 GGUF · unsloth | 22.8 GB | 27.6 GB | 8.0 GB too big | practically identical to the original | Hugging Face → |
| Q6_K GGUF · unsloth | 17.8 GB | 22.6 GB | 3.0 GB too big | very close to the original | Hugging Face → |
| Q5_K_M GGUF · unsloth | 16.1 GB | 20.9 GB | spills 1.3 GB | close; small differences in fine detail | Hugging Face → |
| Q4_K_M ← GGUF · unsloth | 14.3 GB | 19.1 GB | fits · 0.5 GB spare | good; some loss in fine detail and text | Hugging Face → |
| Q3_K_M GGUF · unsloth | 10.8 GB | 15.6 GB | fits · 4.0 GB spare | noticeable loss of detail | Hugging Face → |
| Q2_K GGUF · unsloth | 8.3 GB | 13.1 GB | fits · 6.5 GB spare | heavy loss; a last resort | Hugging Face → |
Sizes from the Hugging Face file listing, read 2026-09-25. “Needed” = file + 4 GB working memory for this model + 0.8 GB kept free for the system.
03The text encoder
Gemma 3 12B: 24.4 GB as 16-bit, 13.2 GB as FP8, 9.4 GB as FP4 mixed. ComfyUI encodes the prompt first and can push the encoder out of VRAM before sampling, so it does not have to fit together with the model. On 19.6 GB the FP8 encoder fits on its own, so prompt encoding stays fast.
04About this GPU
A Mac with 24 GB shares that memory between the CPU and the GPU. On current macOS the GPU may use about 19.6 GB of it by default (older macOS versions: about 17.2 GB); that is the figure used here. ComfyUI can go past it, but macOS then starts compressing and swapping memory and everything slows down. ComfyUI runs on Apple GPUs through PyTorch's MPS backend. 16-bit and GGUF files work; FP8 and INT8 files do not save memory on a Mac, so they are skipped. Speed is the catch: even the fastest Macs are several times slower per image than a desktop RTX card. Chips sold with 24 GB: M2, M3, M4, M4 Pro, M5, M5 Pro, M6 (memory bandwidth 100–307 GB/s — the higher, the faster). Everything about Macs and local AI →
Mac 24 GB: 24 GB unified · ~19.6 GB for the GPU · 100–307 GB/s · Apple Silicon. Everything that runs on the Mac 24 GB →
05What more VRAM would change
With 32 GB you could run LTX-2.3 at 8-bit or better (Q8_0, 22.8 GB) with no offloading: for example on the RTX 5090 32 GB.
06Measured and reported results
Nobody has sent measured numbers for this pair yet. If you run LTX-2.3 on a Mac 24 GB, send your time per image and peak VRAM and it will appear here, credited.
07Questions
How much VRAM does LTX-2.3 need?
Around 13.1 GB with the smallest file (Q2_K) and 27.6 GB with an 8-bit file (Q8_0), counting working memory and a small system reserve. The full 16-bit file needs about 46.8 GB.
Which LTX-2.3 file should I download for the Mac 24 GB?
Q4_K_M (14.3 GB) from unsloth/LTX-2.3-GGUF. Load it with Unet Loader (GGUF) from the ComfyUI-GGUF node pack.
Is FP8 faster than GGUF on the Mac 24 GB?
No. This GPU has no FP8 compute, so ComfyUI converts FP8 weights back before the maths. FP8 only saves memory here; Q8_0 GGUF is the closer-to-original 8-bit pick.
How much do I need to download for LTX-2.3 on the Mac 24 GB?
About 42.8 GB for the model file, text encoder and VAE listed on this page, and the same again free on disk. With a 19.6 GB card, 64 GB of system RAM or more is recommended.