Files read on Hugging Face: 2026-10-24

LTX-2.3 on Mac: files, RAM and published times

LTX-2.3 is a 22B DiT-based audio-video foundation model from Lightricks that generates synchronized video and audio in a single model. On a Mac it runs through community MLX ports, and no official Lightricks Mac benchmark was found.

Lightricks · Updated 2026-10-24
Answer

LTX-2.3 runs on a Mac with 128 GB or more. A published run on M5 Max · 128 GB used the Official FP8 dev checkpoint, quantized to 4-bit at load build (source, 2026-04-11).

Smallest Mac with a published run
128 GBM5 Max · 128 GB · Official FP8 dev checkpoint, quantized to 4-bit at load
Smallest complete download
26.74 GBUnsloth GGUF dev Q4_K_M with GGUF Gemma 3 (ComfyUI)
Files read
2026-10-24Hugging Face

LTX-2 Community License Agreement. Commercial use: restricted. Excluded territories: none listed. Entities with annual revenues of at least $10,000,000 require a paid commercial use license.

See the exact GGUF files and sizes

Table 1

What do you download?

Only builds with every required component and an exact probed file size are shown.

Official FP8 dev checkpoint, quantized to 4-bit at load

safetensors · 4-bit at load from FP8 · 61.13 GB complete · Recommended

Mac routes: appautomaton/ltx-video-mlx

Exact component files for Official FP8 dev checkpoint, quantized to 4-bit at load
FileComponentRequiredDownloadSource
ltx-2.3-22b-dev-fp8.safetensorsLightricks/LTX-2.3-fp8LTX-2.3 22B transformer (dev or distilled)ditYes29.15 GBHugging FaceRead 2026-10-24
ltx-2.3-22b-distilled-lora-384.safetensorsLightricks/LTX-2.3Distilled LoRAotherNo7.61 GBHugging FaceRead 2026-10-24
model-00001-of-00005.safetensorsgoogle/gemma-3-12b-ptGemma 3 12B text encodertext_encoderYes4.98 GBHugging FaceRead 2026-10-24
model-00002-of-00005.safetensorsgoogle/gemma-3-12b-ptGemma 3 12B text encodertext_encoderYes4.93 GBHugging FaceRead 2026-10-24
model-00003-of-00005.safetensorsgoogle/gemma-3-12b-ptGemma 3 12B text encodertext_encoderYes4.93 GBHugging FaceRead 2026-10-24
model-00004-of-00005.safetensorsgoogle/gemma-3-12b-ptGemma 3 12B text encodertext_encoderYes4.93 GBHugging FaceRead 2026-10-24
model-00005-of-00005.safetensorsgoogle/gemma-3-12b-ptGemma 3 12B text encodertext_encoderYes4.60 GBHugging FaceRead 2026-10-24

Build sources: appautomaton ltx-video-mlx README.

Unsloth GGUF dev Q4_K_M with GGUF Gemma 3 (ComfyUI)

gguf · Q4_K_M · 26.74 GB complete

Mac routes: ComfyUI

Exact component files for Unsloth GGUF dev Q4_K_M with GGUF Gemma 3 (ComfyUI)
FileComponentRequiredDownloadSource
ltx-2.3-22b-dev-Q4_K_M.ggufunsloth/LTX-2.3-GGUFLTX-2.3 22B transformer (dev or distilled)ditYes14.33 GBHugging FaceRead 2026-10-24
vae/ltx-2.3-22b-dev_video_vae.safetensorsunsloth/LTX-2.3-GGUFVideo VAE (bundled in the official checkpoints)vaeNo1.45 GBHugging FaceRead 2026-10-24
vae/ltx-2.3-22b-dev_audio_vae.safetensorsunsloth/LTX-2.3-GGUFAudio VAE (bundled in the official checkpoints)audioNo0.36 GBHugging FaceRead 2026-10-24
text_encoders/ltx-2.3-22b-dev_embeddings_connectors.safetensorsunsloth/LTX-2.3-GGUFText projection and connectors (bundled in the official checkpoints)otherNo2.31 GBHugging FaceRead 2026-10-24
gemma-3-12b-it-qat-UD-Q4_K_XL.ggufunsloth/gemma-3-12b-it-qat-GGUFGemma 3 12B text encodertext_encoderYes7.43 GBHugging FaceRead 2026-10-24
mmproj-BF16.ggufunsloth/gemma-3-12b-it-qat-GGUFGemma 3 12B text encodertext_encoderYes0.85 GBHugging FaceRead 2026-10-24

Build sources: Unsloth LTX-2.3 GGUF card.

Official dev checkpoint (BF16) with Gemma 3

safetensors · BF16 · 70.52 GB complete

Mac routes: Lightricks/LTX-2 ltx-pipelines

Exact component files for Official dev checkpoint (BF16) with Gemma 3
FileComponentRequiredDownloadSource
ltx-2.3-22b-dev.safetensorsLightricks/LTX-2.3LTX-2.3 22B transformer (dev or distilled)ditYes46.15 GBHugging FaceRead 2026-10-24
model-00001-of-00005.safetensorsgoogle/gemma-3-12b-it-qat-q4_0-unquantizedGemma 3 12B text encodertext_encoderYes4.98 GBHugging FaceRead 2026-10-24
model-00002-of-00005.safetensorsgoogle/gemma-3-12b-it-qat-q4_0-unquantizedGemma 3 12B text encodertext_encoderYes4.93 GBHugging FaceRead 2026-10-24
model-00003-of-00005.safetensorsgoogle/gemma-3-12b-it-qat-q4_0-unquantizedGemma 3 12B text encodertext_encoderYes4.93 GBHugging FaceRead 2026-10-24
model-00004-of-00005.safetensorsgoogle/gemma-3-12b-it-qat-q4_0-unquantizedGemma 3 12B text encodertext_encoderYes4.93 GBHugging FaceRead 2026-10-24
model-00005-of-00005.safetensorsgoogle/gemma-3-12b-it-qat-q4_0-unquantizedGemma 3 12B text encodertext_encoderYes4.60 GBHugging FaceRead 2026-10-24

Build sources: LTX-2.3 models list.

Table 2

Does it fit your Mac's RAM?

Published names the exact measured machine. est. is a memory calculation. unknown means the run verdict is not established, not that the model fails.

Run verdicts by Mac memory tier
Mac RAMComplete buildDownloadRun verdictBasis
8 GBOfficial FP8 dev checkpoint, quantized to 4-bit at load4-bit at load from FP8 · appautomaton/ltx-video-mlx61.13 GBDoes not fitest.
16 GBOfficial FP8 dev checkpoint, quantized to 4-bit at load4-bit at load from FP8 · appautomaton/ltx-video-mlx61.13 GBDoes not fitest.
24 GBOfficial FP8 dev checkpoint, quantized to 4-bit at load4-bit at load from FP8 · appautomaton/ltx-video-mlx61.13 GBDoes not fitest.
32 GBUnsloth GGUF dev Q4_K_M with GGUF Gemma 3 (ComfyUI)Q4_K_M · ComfyUI26.74 GBunknownunknown
48 GBUnsloth GGUF dev Q4_K_M with GGUF Gemma 3 (ComfyUI)Q4_K_M · ComfyUI26.74 GBunknownunknown
64 GBOfficial FP8 dev checkpoint, quantized to 4-bit at load4-bit at load from FP8 · appautomaton/ltx-video-mlx61.13 GBunknownunknown
96 GBOfficial FP8 dev checkpoint, quantized to 4-bit at load4-bit at load from FP8 · appautomaton/ltx-video-mlx61.13 GBunknownunknown
128 GBOfficial FP8 dev checkpoint, quantized to 4-bit at load4-bit at load from FP8 · appautomaton/ltx-video-mlx61.13 GBRuns · PublishedPublishedApple M5 Max · 128 GB RAM

How does it rank in blind votes?

No independent score yet.

Table 3

What published Mac times are available?

Third-party results keep their original machine and settings. ModelFit does not run these benchmarks.

Published Mac generation measurements
Machine and RAMSettingsAppResultSource
Apple M5 Max · 128 GB RAM512x512 · 4.8 s · 8 steps · 121 frames · 4-bit at load from FP8 · T2V, end-to-end total with approximate figures (the source marks them with ~). The repository carries no run date, so the date shown is the pinned commit date.appautomaton/ltx-video-mlxSuccess~101s~53 GBappautomaton ltx-video-mlx performance notesPublished 2026-04-11 · Grade A

Which Mac software routes are listed?

  • ComfyUI

    Official

    Formats: safetensors, gguf. Memory policy: component stages stay resident.

    On Apple Silicon ComfyUI runs on PyTorch MPS, not MLX. GGUF files need the separate city96/ComfyUI-GGUF custom node. Resident size is the sum of the weights, an upper bound: ComfyUI can unload models between stages and publishes no per-model Mac peak.

    Sources: ComfyUI README · ComfyUI system requirements · ComfyUI CLI arguments · ComfyUI-GGUF README

  • Lightricks/LTX-2 ltx-pipelines

    Official

    Formats: safetensors. Memory policy: component stages stay resident.

    No Lightricks Mac benchmark was found.

    Sources: LTX-2 README

  • appautomaton/ltx-video-mlx

    Community

    Formats: safetensors. Memory policy: component stages stay resident.

    Sources: appautomaton ltx-video-mlx README

What does the official configuration say?

"Official tested stack: Python >=3.12, CUDA version >12.7, PyTorch ~= 2.7. No official Mac memory figure was found."

Source: LTX-2.3 model card · Read 2026-10-24.

What known issues should you check?

  • Known issue 1

    The card lists only a CUDA test stack, and no official Mac memory statement or Lightricks Mac benchmark was found.

    Source: LTX-2.3 model card · Read 2026-10-24

  • Known issue 2

    The appautomaton benchmark carries no run date. The date shown is the date of the pinned commit.

    Source: appautomaton ltx-video-mlx performance notes · Read 2026-10-24

  • Known issue 3

    Width and height must be divisible by 32, and the frame count must be divisible by 8 plus 1.

    Source: LTX-2.3 model card · Read 2026-10-24

Frequently asked questions

How fast is LTX-2.3 on a Mac?

appautomaton's MLX port generated 121 frames at 512x512 in about 101 s end to end on an M5 Max with 128 GB, using 4-bit weights and 8 steps, with roughly 53 GB peak RAM. The source gives no run date, so the date shown is the commit date.

What does the LTX-2.3 download include?

Each main checkpoint is one file bundling the transformer, the video and audio VAEs and the text projection. The Gemma 3 text encoder is downloaded separately, and width and height must be divisible by 32.

Sources