LTX

2 aliases · LTX-2.5 (flagship) · LTX-2.3. Lightricks' video line, MLX-native — the only local backend whose clips come back with synchronized audio.

LTX-2.5 is the flagship of rapid-mlx's video lane: text-to-video and image-to-video through the OpenAI-compatible /v1/videos jobs API, with the finest size control of any backend (any 32-pixel-aligned resolution from 256 to 1920) and a settable frame rate up to 60 fps — and it generates synchronized audio, which rapid-mlx preserves in the MP4. The checkpoint is a 67.7 GB download but serves on a 24 GB Mac: it uses the low-RAM distilled path by default.

The two lines split cleanly: 2.5 preserves synchronized audio; 2.3's silent track is remuxed away by design — details in each section below.

family

LTX (Lightricks)

aliases

2

audio

synchronized, preserved (2.5)

install

pip install 'rapid-mlx[video]'

also needs

ffmpeg · Python 3.11+

OpenAI base URL

http://localhost:8000/v1

Usage · ltx-2.5

pip install 'rapid-mlx[video]'
brew install ffmpeg
rapid-mlx serve ltx-2.5-mlx-q8

# Text-to-video — the MP4 comes back with synchronized audio
curl http://127.0.0.1:8000/v1/videos \
  -F model=ltx-2.5-mlx-q8 \
  -F 'prompt=a fox running through fresh snow, cinematic tracking shot' \
  -F seconds=4 \
  -F size=1280x720 \
  -F seed=42

# Image-to-video: animate a still, audio included
curl http://127.0.0.1:8000/v1/videos \
  -F model=ltx-2.5-mlx-q8 \
  -F 'prompt=a paper boat drifting down a rain gutter, macro, shallow focus' \
  -F input_reference=@still.png \
  -F conditioning_strength=0.8 \
  -F seconds=4 \
  -F size=1280x720

# Poll, then download
curl http://127.0.0.1:8000/v1/videos/video_ID
curl http://127.0.0.1:8000/v1/videos/video_ID/content --output result.mp4

Download

Both checkpoints are mirrored on the rapid-mlx CDN — our fresh-install test measured pulls at 90 MB/s, and if a mirror file ever slows down, the pull finishes that file from Hugging Face automatically, mid-pull, no restart. One command each, no account:

LTX-2.5 · 67.7 GB · 23 files · mirrored ✓ · live status

LTX-2.3 · 22.8 GB · 15 files · mirrored ✓ · live status

Weights land in the standard Hugging Face cache, so other MLX tools can reuse them — and rapid-mlx serve pulls automatically on first use, so pull is just doing it ahead of time. Prefer the source? Both repos are linked in the alias table below.

Lines in this family

• LTX-2.5 · LTX-2.3

LTX-2.5 · 1 alias · flagship

The current line, and the reason this family leads the video lane: it generates synchronized audio, and rapid-mlx preserves that track in the MP4 — the only local backend whose clips come back with sound. A 67.7 GB download that serves on a 24 GB Mac via its low-RAM distilled path.

What LTX-2.5 accepts

• Size — any 32-pixel-aligned width × height from 256 to 1920; 1280x720 and 720x1280 as presets. The 32-px granularity is the finest of the three backends (Wan and CogVideoX-Fun align to 64).
• Frames — 8n+1, from 9 to 1201; seconds (1–20, default 4) derives the count if you don't pass one.
• fps — settable, 1 to 60. (Wan, by contrast, is fixed to its checkpoint's native rate.)
• Deliberately minimal knobs — guidance_scale and negative_prompt are rejected rather than ignored, so a request that would silently do nothing fails loudly instead. conditioning_strength (0–1) on image-to-video is the one dial.

Ask the running server rather than memorising any of this — GET /v1/videos/capabilities reports the live model's modes, size rule, frame rule and fps bounds, the same contract the route validates against.

LTX-2.3 · 1 alias

The previous line, kept registered. Its checkpoint is an audio-video one whose audio track is silent, so since 0.11.4 rapid-mlx remuxes it away by design and the MP4 comes back video-only — a silent track is worse than none, because downstream tools read it as "this clip has sound". Standard 8n+1 frame rule and conditioning_strength on image-to-video, like 2.5; unlike 2.5 it still accepts guidance_scale and negative_prompt. At a third of the download it remains the quick way to try the family on limited disk.

Notes & caveats

• Jobs are serialized — a second request queues rather than running concurrently. Two diffusion pipelines resident at once would exhaust unified memory.
• A clip is minutes of compute, not milliseconds. Lowering diffusion steps is the biggest speed lever, at a quality cost.
• Image-to-video takes a real uploaded file; remote image URLs are refused rather than fetched (SSRF).

Frequently asked questions

Can I generate video with audio locally on a Mac?

Yes. LTX-2.5 generates synchronized audio and rapid-mlx preserves that track in the MP4 — the only local video backend on rapid-mlx whose clips come back with sound. Serve ltx-2.5-mlx-q8 and POST /v1/videos; everything renders on your own machine.

How much RAM does LTX-2.5 need on a Mac?

24 GB of unified memory. The checkpoint is a 67.7 GB download, but rapid-mlx uses its low-RAM distilled path by default, so it serves on a 24 GB Mac. LTX-2.3 q4 (~22.8 GB download) also runs at 24 GB.

What controls does LTX-2.5 support?

Any 32-pixel-aligned resolution from 256 to 1920, frame counts of 8n+1 up to 1201, fps up to 60, and conditioning_strength for image-to-video. guidance_scale and negative_prompt are deliberately rejected rather than ignored.

Where next