From the next Kynd update, 16 GB Apple Silicon Macs run LTX 2.5's one-stage tiers locally — draft, balanced, and the experimental Ingredients route at its official 768×448 geometry — through a new streamed engine with a measured 13.3 GB peak for a five-second clip. The 32 GB resident route is unchanged, byte for byte.
Local video's hard wall has always been unified memory. Metal doesn't swap; it aborts. So when LTX 2.5's engine held its text encoder, connector and 22-billion-parameter transformer resident at once, the measured 28 GB peak made 32 GB an honest floor. The fix wasn't a smaller model. It was refusing to hold everything at once.
The memory ladder, measured.
Three changes, in plain language.
Never two giants at once
- The 12 B text encoder runs while the transformer slot is empty.
- Embeddings hand off; the encoder is freed.
- Then the transformer loads for the denoise.
- Repeat prompts skip the encoder entirely via a small on-disk embedding cache.
One block resident, not 48
- The 22 B transformer's weights are read lazily per denoising step.
- Each of the 48 blocks materialises, computes, and is freed.
- ~11 GB of resident weights becomes roughly one block at a time.
- Verified byte-identical to the resident engine at two clip lengths.
Half a frame at a time
- Turning latents into pixels was the single largest spike.
- The frame decodes as two overlapping halves.
- A 128-px cross-fade joins them — measured 45.6 dB PSNR against untiled.
- The seam runs through faces without a trace.
Proof, not promises.
Streaming changes when weights occupy memory, never what the model computes — so the streamed engine's raw output was checksummed against the resident engine's on identical seeds, at both two and five seconds. The checksums match exactly. Tiled decode blends a seam, so it is held to a different standard: 45.6 dB PSNR against the untiled frame, with the worst frame at 45.1 dB — comfortably beyond the threshold where differences stop being visible. It ships off by default; only low-memory tiers opt in.
Speed held too. Streaming costs between 1.5 and 11 percent per render, and repeated prompts come out faster than the old engine, because the embedding cache skips the multi-gigabyte text encoder pass entirely.
What a 16 GB Mac gets.
16 GB LTX 2.5 FAQ.
Can a 16 GB MacBook Air really run LTX 2.5?
Yes, from the next Kynd update, on the one-stage tiers through the streamed engine. Kynd unloads chat models before video renders, and the tier is labelled experimental while the 13.3 GB envelope gathers real-world mileage.
Is the quality reduced to make it fit?
No. Same model, same weights, same steps. The streamed denoise is checksum-identical to the resident engine; tiled decode measures 45.6 dB PSNR against the untiled path, which is visually indistinguishable.
What about the compact LTX 2.3 route?
It remains available and is still the lightest option. But 16 GB Macs no longer trade down to 2.3 for reference-conditioned work — the Ingredients route that produces consistent characters is a 2.5 capability, and it now fits.
Does this change 32 GB and larger Macs?
Their default route is byte-unchanged. They also gain the prompt-embedding cache, which makes repeat renders start faster.