LTX-2.5 on Mac: Memory, Files and Published Timings (2026)

LTX-2.5 runs on a Mac through community MLX ports, and every published Mac run in the sources checked on 2026-10-24 sits at 128 GB or on a chip with no memory size stated. An M3 Max with 128 GiB took 619.57 seconds to render 4.042 seconds of 704x480 video (vanch007 benchmark, 2026-08-12). The official quick start downloads roughly 66 GiB in per-component files, with a fine-tuned Gemma 4 12B text encoder where LTX-2.3 used Gemma 3 (Lightricks README, 2026). Below are the files, the Mac routes and each dated timing, with every LTX-2.3 run labeled so the two versions stay apart.

Which Macs can run LTX-2.5?

Macs with 128 GB have published runs. The 64 GB verdict remains unknown: the default bf16 layout peaked at 62.40 GB on a 128 GB machine for a test clip, but that does not prove a tight or successful fit on a 64 GB Mac. Macs with 32 GB or less have a documented low-memory mode and no recorded LTX-2.5 run.

The mlx-community peaks come from a 128 GB Mac running a Swift consumer. The vanch007 footprint is from an M3 Max with 128 GiB, the appautomaton figure from an M5 Max with 128 GB, and the xocialize figure from an M5 Pro with 24 GB. vanch007 reports GiB and the others report GB.

The mlx-community author calls the text encoder the surprise. It is an unquantized bf16 Gemma 4 12B, 23.8 GB on disk and about 24.4 GB in memory, where LTX-2.3 packs used a 4-bit Gemma 3 of about 7 GB (mlx-community, 2026-08-13). The DiT adds a 37.98 GB resident floor in bf16, and the two together come close to the 62.40 GB default peak. The author also warns that numbers change with resolution and frame count, and 448x320 at 9 frames is a test clip.

Lightricks lists NVIDIA hardware for the model: 32GB+ VRAM, 32GB of system RAM and 100GB of free storage, plus a separate Min VRAM label of 16GB (ltx.io, 2026). The same page says LTX-2.5 runs on consumer hardware including Mac and gives no Mac figure.

For 32 GB and 16 GB Macs, the dgrauet port documents a --low-ram mode that streams transformer blocks from disk, so that q8 fits 16 GB Macs and bf16 fits 32 GB Macs (dgrauet README, 2026). Its 16 GB example uses the LTX-2.3 pack, and the README lists no LTX-2.5 timing for either size. The only published low-memory run is LTX-2.3 on a 24 GB M5 Pro, in the next section.

How long does an LTX-2.5 clip take on a Mac?

About ten minutes for a four-second clip at 704x480 on an M3 Max with 128 GiB: 619.57 seconds at 8 steps, one run (vanch007 benchmark, 2026-08-12). The three smaller cases finish in 29.77 to 124.50 seconds, but at thumbnail sizes up to 384x256. Each row stays on its own line because time does not scale in a straight line with steps or frames.

All 11 cases of the benchmark ran in fresh processes on a Mac15,8 that the project labels an M3 Max with 128 GiB, between 20:19 and 20:42 (UTC+8) on 2026-08-12, one sample per case. Wall time includes prompt encoding, shard loading, denoising, VAE and audio decoding, and MP4 muxing. The 704x480 run is 10.33 minutes, or 153.3 times the clip length, and peak footprint stayed between 17.00 and 17.06 GiB with 0 swaps.

The benchmark names its pack LTX-2.5-local-v4 and does not state the quantization. The repository's current v5 pack uses a Q8 transformer and a Q4 Gemma 4, and that description is not applied to these runs (vanch007 README, 2026).

The dgrauet README adds one more LTX-2.5 run. On an M1 Max with the 2.5 q8 pack, a 576×1024×241 distilled image-to-video render took 701 seconds by default and 563 seconds with LTX2_COMPUTE_DTYPE=float16 (dgrauet README, 2026). The line gives no RAM size and no date, so it cannot be set against the M3 Max rows.

For scale, Lightricks states that two GB200 GPUs generate a 10-second 720p clip in 6.8 seconds (ltx.io, 2026). That is datacenter hardware and says nothing about a Mac.

The runs below are LTX-2.3. They are labeled because LTX-2.3 uses Gemma 3 and a different pack layout, so neither the memory nor the time carries over to LTX-2.5.

The appautomaton document is dated by its commit, 2026-04-11, and its README names the model LTX-2.3. Its times are rounded with a tilde, and 121 frames is 4.8 seconds in that table. The xocialize README describes a port of LTX-2.3 with a Gemma 3 encoder, although another paragraph mentions LTX-2.5 profiles, so treat the 24 GB rows as LTX-2.3 evidence at best.

What do you download for LTX-2.5?

The official quick start needs five files, about 71 GB in total: the distilled DiT, the Gemma 4 encoder, two VAEs and the spatial upscaler. Lightricks puts that set at roughly 66 GiB (Lightricks README, 2026). Sizes below are rounded from the byte counts the Hugging Face API reports for the Lightricks/LTX-2.5 repository, where the files sit in the diffusion_models, text_encoders, vae and latent_upscale_models folders.

Other files in the same repository are optional. The dev DiT ltx-2.5-22b-dev-transformer-bf16.safetensors is 42 GB and serves the guided two-stage pipelines. The lighter convolutional decoder ltx-2.5-video-vae-conv-bf16.safetensors is 1.5 GB, the temporal upscaler ltx-2.5-latent-temporal-upscaler-x2-bf16-1.0.safetensors is 0.3 GB, and the distilled LoRA ltx-2.5-22b-distilled-lora-450-bf16.safetensors is 8.9 GB. The small ltx-2.5-duration-head-bf16.safetensors (4 MB) lets you omit the frame count and have the clip length predicted from the prompt.

The text encoder is not interchangeable with Google's stock Gemma 4. Loading checks its version against gemma4-12b-ltx-v1, the one the checkpoint was trained with.

The MLX packs repackage these files. The dgrauet repositories hold both transformer-dev.safetensors and transformer-distilled.safetensors, a text_encoder.safetensors carrying the Gemma 4 tower, a connector, and the VAE, vocoder and upscaler files. The README sizes the packs at about 75 GB and about 47 GB because the download holds both transformers, which is not what one pipeline loads (dgrauet README, 2026).

The README also lists a bf16 pack at about 120 GB that runs with --low-ram on 32 GB. The Q8 and Q4 repositories are gated with automatic approval on Hugging Face, so accept the terms and log in before the download (Hugging Face API, 2026).

Which software routes work on a Mac?

Seven routes appear in the sources. The table says which model each one documents and what a Mac user can expect from it.

The official README documents one macOS detail. The natten extra for the video VAE is Linux and CUDA only, so on macOS it is skipped and decoding falls back to a Triton or eager implementation. For GPU memory limits it suggests --quantization fp8-cast with offload, and it gives no Mac recipe for either.

The dgrauet port has the most Mac features. It detects LTX-2.5 from the pack, so the command needs no version flag. On 2.5 packs -f is optional because the DurationHead predicts the clip length, while on 2.3 packs -f stays required.

git clone https://github.com/dgrauet/ltx-2-mlx.git
cd ltx-2-mlx
uv sync --all-extras
ltx-2-mlx generate --distilled --model dgrauet/ltx-2.5-mlx-q8 \
  --prompt "a heavy wooden door creaks slowly open" --frame-rate 24 -o out.mp4

The README shows --model /path/to/ltx-2.5-mlx-q8 and says a Hugging Face repo id works too. Add --low-ram to stream blocks from disk. The vanch007 repository calls itself a research workspace, so follow its README step by step. For the Mac side of ComfyUI, see ComfyUI on Mac.

Which GGUF file should you pick for LTX-2.5?

Unsloth recommends UD-Q4_K_XL at 14.0 GB for most GPUs and UD-Q2_K_XL at 9.5 GB for small cards (Unsloth, 2026). The GGUF replaces only the DiT, so ComfyUI still needs the official text encoder, VAEs and latent upscaler. Unsloth documents ComfyUI-GGUF and its own Studio, names no Mac, and no Mac timing for these files appears in the sources checked.

LPIPS and PSNR compare each file's render with the bf16 DiT on the same prompt and seed. Lower LPIPS and higher PSNR mean closer to bf16. Unsloth ran distilled clips at 1280x704 with 121 frames, 8 + 3 steps and CFG 1, over four seeds.

A full ComfyUI set adds about 29 GB of support files: the 26 GB bf16 encoder, the 1.5 GB video VAE, the 0.4 GB audio VAE and the 1.0 GB spatial upscaler. So UD-Q4_K_XL ends near 43 GB on disk. Dev files sit at the repository root and distilled files in distilled/. Distilled runs 8 + 3 steps at CFG 1, and dev runs 30 steps at CFG 3.

LTX-2.3 has its own set, unsloth/LTX-2.3-GGUF, with dev, distilled and distilled-1.1 folders. Its distilled Q4_K_M is 14 GB. It pairs with a Gemma 3 GGUF encoder, gemma-3-12b-it-qat-UD-Q4_K_XL.gguf plus mmproj-BF16.gguf, where the 2.5 set takes the Gemma 4 encoder from Lightricks. Do not mix the sets, because Lightricks says the files are not interchangeable.

What changes between LTX-2.3 and LTX-2.5?

Four things matter on a Mac: how the files are packaged, which text encoder loads, how much memory that encoder takes, and which license applies. Both versions have a 22B DiT. Files and LoRAs do not carry across (Lightricks README, 2026).

The memory row matters most. A setup sized from LTX-2.3 numbers under-estimates LTX-2.5, as the mlx-community author says. The Q8 pack grows from about 26 GB to about 75 GB, though the README adds that the 2.5 size is the download and not what one pipeline loads. The dgrauet prompt enhancer is Gemma 3 only, so enhance raises an error on 2.5 packs.

Both licenses draw the same line. Entities with annual revenues of at least $10,000,000 need a paid license for use beyond a narrow non-commercial purpose (LTX-2.x license, 2026).

Which setup should you start with?

Going only on the published evidence, by Mac memory:

• At 128 GB, use dgrauet's Q8 pack or the vanch007 port. The one published timing is about ten minutes for 4 seconds at 704x480 on an M3 Max.
• At 64 GB, the verdict stays unknown because no timing exists. In the mlx-community test clip, evicting the DiT around encoding took the peak from 62.40 GB to 40.66 GB on a 128 GB machine.
• At 32 GB and under, LTX-2.5 has a documented --low-ram path and no result. LTX-2.3 has the only low-memory timing, 150 seconds for a 704x512, 121-frame clip on a 24 GB M5 Pro.

The LTX-2.3 model page lists its complete builds with a RAM verdict per Mac tier. For other video models and Mac tiers, start at the video hub. The local video generation guide for Mac covers the method, and the best Mac for AI video generation covers hardware choices. For text models, the homepage configurator picks by Mac.

FAQ

Can LTX-2.5 run on a 16 GB Mac?

No published run shows it. The dgrauet README documents a --low-ram mode and says q8 fits 16 GB Macs, but its 16 GB example uses the LTX-2.3 pack. The xocialize port calls 16 GB deliberately unsupported for its LTX-2.3 profiles. Treat 16 GB as untested for LTX-2.5 until someone publishes a run with memory and settings.

Is LTX-2.5 free for commercial use?

Under the LTX-2.x Community License dated 2026-08-11, entities with annual revenues of at least $10,000,000 need a paid license for any use beyond a narrow non-commercial purpose. Below that line the license is royalty-free, subject to its acceptable-use attachment. Read the license text before shipping anything, since this page is not legal advice.

Can I use LTX-2.3 files or LoRAs with LTX-2.5?

No. Lightricks says the files are not interchangeable between the two models and that a LoRA only works with the model it was trained on. LTX-2.3 bundles the transformer, VAEs and projection in one checkpoint, while LTX-2.5 splits them per component, so download the 2.5 files in full.

Does LTX-2.5 make audio on a Mac?

Yes, in the MLX ports that cover it. LTX-2.5 is a joint audio-video model, and the dgrauet README says its pipeline generates video plus stereo 48kHz audio. The vanch007 timings include the audio decode and the MP4 muxing, so the 619.57 seconds covers sound as well as picture.

How long will a five-second clip take?

No published run answers that for a typical Mac. The closest is 619.57 seconds for 4.042 seconds of 704x480 video on an M3 Max with 128 GiB, one run at 8 steps. Chip, pack, size and steps all move the number, so read it as an order of magnitude and never as a forecast.

Sources

Read on 2026-10-24 unless a date is given.

• Lightricks, LTX-2 README: quick start, component list, LTX-2.3 note.
• Lightricks, LTX-2.3 model list and LTX-2.3 model card.
• Lightricks, LTX-2.x Community License (2026-08-11) and LTX-2 Community License (2026-01-05).
• LTX, LTX-2.5 model page: hardware requirement, Mac statement, GB200 figure.
• Hugging Face API file sizes: Lightricks/LTX-2.5, Lightricks/LTX-2.3, dgrauet/ltx-2.5-mlx-q4 and unsloth/LTX-2.3-GGUF.
• vanch007, M3 Max benchmark (run 2026-08-12) and repository README.
• dgrauet, ltx-2-mlx README: requirements, packs, the M1 Max run.
• mlx-community, ltx-2.5-mlx model card (status 2026-08-13): memory peaks.
• Unsloth, LTX-2.5-GGUF: file sizes and LPIPS against bf16.
• appautomaton, performance document at commit 83b4881 (2026-04-11).
• xocialize, ltx-2-mlx-swift README: compact24 profile on a 24 GB M5 Pro.
• Blaizzy, mlx-video README: documents LTX-2 and LTX-2.3.