LTX 2.5 22B quantized to Q4 runs text-to-video on an 8GB M1 MacBook Air
iJohnnyRobert · reddit · 2026-10-19
A Reddit user quantized LTX 2.5 distilled 22B to Q4 and got it running locally on an M1 MacBook Air with just 8GB unified memory, generating 10 seconds of 256×160 video with 48kHz stereo audio in 110 seconds — pure text-to-video.
Key technical details:
- Experimental MLX/Metal setup: streamed model blocks, sequential text encoding, staged unloading, tiled decoding
- 6 sampling steps; 12fps generation interpolated to 24fps
- Initial Metal allocation peaked at 7.12 GiB (7.64GB); after optimizing decoding, a later run peaked at only 3.03 GiB, finishing in 2m43s
Visuals are still rough drafts, but the audio quality was surprisingly good and system memory never felt strained — the author expected the whole laptop to freeze.
More from Infra
- Google and Synaptics Preview Coral Microcontroller for On-Device AI at Under 6 Milliwatts — lmoroney · 2026-10-19
- Not Paris looters — just Series D founders who still can't afford GPUs — DesireeCachette · 2026-10-19
- Sovereign AI Blog ships an MCP server for its NVIDIA DGX Spark self-hosting log — modelcontextprotocol · 2026-10-19
- Why agent sandboxes became a thing: slow VMs pushed Lovable, E2B and Modal to build their own — vboykis · 2026-10-19
- Local Qwen vision model builds a 3D room game in 26 minutes on a single prompt — Beneficial-Ad-8127 · 2026-10-19
- RTX 3090 runs 27B LLM 4.6x faster via pure inference optimizations, no new hardware — rickasaurus · 2026-10-19