LTX 2.5 now fits a 16 GB Mac.

MLX video · Apple Silicon

Same pixels. Half the memory.

Three engine changes. One measured 13.3 GB peak.

Actual output of the 13.3 GB render

Quick answer

Kynd’s tested low-memory route lets 16 GB Apple Silicon Macs run LTX 2.5’s one-stage tiers locally — Social, draft, balanced, and the Ingredients route at its official 768×448 geometry. Streaming is now the default on every supported memory tier because it measured both smaller and faster than keeping the transformer resident.

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.

Every number on this page is a measured Metal peak from the engine’s own accounting — not an estimate.

Shorter clips go lower. A two-second Ingredients take peaks at a measured 9.4 GB. Plain text-to-video tiers sit below the Ingredients numbers because they carry no doubled reference token stream.

Three changes, in plain language.

1 · Staged phases

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.

2 · Block streaming

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.

3 · Tiled decode

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 now runs by default on every one-stage tier.

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.

Labelled honestly. The Social route is proven from 16 GB upward and ships with a conservative 13 GB envelope. Ingredients remains a separate experimental capability at 16 GB. Macs with 24 GB or more run the same streamed engine with more headroom, not a different quality recipe.

16 GB LTX 2.5 FAQ.

Can a 16 GB MacBook Air really run LTX 2.5?

Yes. The one-stage tiers run through the streamed engine. Kynd unloads chat models before video renders; the proven Social route uses a conservative 13 GB envelope, while Ingredients remains separately labelled experimental at 16 GB.

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.

Also useful: The full LTX-on-Mac tier guide → · The LTX 2.5 Ingredients launch →

The wall was residency, not the Mac.

Kynd measures the peak, gates the tier, and streams what never needed to sit in memory whole.

Latest local render

Twenty seconds. Four camera setups. One take.

Three inputs condition this render at once — a reference frame for the character, the real vocal stem frozen onto the timeline, and a prompt that names four separate camera setups. LTX 2.5 rendered all 481 frames as one continuous local take, and the song runs unbroken across every change of angle. Sound on.

The camera never travels between those positions. It is repositioned, so the picture changes instantly — while the character, the wardrobe, the van and the light hold across every change, and the mouth stays on the words in all four setups.

Social works from 16 GB upward. The current one-stage Fast Path streams transformer blocks and tile-decodes the result instead of holding the whole pipeline resident. A full-length 481-frame streamed validation peaked at 9.15 GB; Kynd ships the Social tier with a conservative 13 GB envelope and a 16 GB gate. The same model, steps and creative recipe run on 16, 24, 32 and 64 GB Macs.

Rendered with this exact pipeline

He’s singing the actual song. On a Mac.

One character reference image, one music track, one request: Kynd’s Video Studio locks the real vocal to the timeline, so the mouth follows the words — not an impression of them. Sound on.