Instructions to use xocialize/ltx-2.5-granules with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use xocialize/ltx-2.5-granules with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir ltx-2.5-granules xocialize/ltx-2.5-granules
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
LTX-2.5 granules β per-block weight streaming (MLX)
A re-layout of LTX-2.5's DiT into per-block granule files, so the transformer can be streamed from SSD during denoise instead of held resident. Same bytes, reorganised.
| tree | source | streamed |
|---|---|---|
bf16/ |
mlx-community/ltx-2.5-mlx |
34.58 GiB |
q8/ |
mlx-community/ltx-2.5-mlx-ditq8 |
18.38 GiB |
Consumed by ltx-2-mlx-swift via
BlockStreamKit. Sibling of
xocialize/ltx-2.3-granules.
Fetch these INSTEAD of the transformer, not as well
Each tree carries a globals.granule sidecar holding the 59 non-block tensors, so a consumer binds
without the source checkpoint present at all. Downloading both would double the DiT's disk for
no benefit.
Output is bit-identical, not merely close
Streamed vs resident is a memcmp gate, not a cosine gate β paired with a poisoned-slot negative control proving the comparison can fail. Quality is not a consideration when choosing this path.
What it buys, and what it costs
Buys: the resident DiT floor disappears (slots are ~1.5 GiB for q8, ~2.9 GiB for bf16). That is what makes LTX-2.5 fit 24 GB and 32 GB machines at all, and on 64 GB it measured faster than resident β eviction otherwise reloads the whole DiT every stage.
Costs: ~6Γ read amplification (a full sweep per denoise step rather than one load). Overlapped, so no wall-clock penalty at real generation sizes β but real SSD and power load, and unmeasured on battery.
β οΈ Streaming is only profitable above a token threshold the runtime computes from measured IO vs compute. Below it the consumer falls back to resident output-invisibly. On small-N tiers that estimate is noisy enough to flip run to run, so the consumer pins the decision there rather than letting it vary.
Provenance
Each manifest.json records source_repo + source_revision alongside per-file SHA-256. Those
stamps are verified at layout time against the live Hub β the declared revision must resolve and
the declared file size must match β so a downloader can check what a tree came from rather than
inferring it from a filename.
Licence
LTX-2 Community License β see LICENSE.md and the acceptable-use snapshot. These are a derivative
of Lightricks LTX-2.5 and carry its terms, including the Attachment A revenue gate and non-compete.
Redistribution permitted under Β§3 with those terms propagated.
Model tree for xocialize/ltx-2.5-granules
Base model
Lightricks/LTX-2.5