Three Lite Voice AI Apps Are Live: They Run on NVIDIA, AMD and Intel GPUs—and on Macs—from 2GB VRAM ​

We · October 2026

This is the first article in the "Product Updates" series: three lite voice-synthesis apps are now live in our catalog. If you want the mechanics of what quantization actually removes, start with Parameters, Quantization, VRAM: Why “It Fits” Is Not “It Runs” in the "AI Tech Explainers" series.

AI moves fast. The features and requirements described here reflect where the products stand right now, and are for reference only—each app's own page stays authoritative.

The short version: what shipped, and who can use it ​

Three voice-synthesis (TTS) apps now have lite editions in the catalog:

Shared baseline for all three: at least 16GB of system memory and about 10GB of free disk. Windows 10/11 and macOS 15+ are both supported; the Mac branches are spelled out in the next section.

Same-machine test: the lite edition is faster. One computer, one 1,153-character article, run once on each edition of IndexTTS 2.5 (RTF = synthesis time ÷ audio length; lower is faster):

Which GPUs are supported: the plain list ​

Hardware range is the biggest difference from the originals. The originals accelerate through CUDA, which only NVIDIA cards speak, and the IndexTTS 2 / 2.5 originals demand 8GB of VRAM. The lite editions move to the Vulkan backend, which widens support to:

• NVIDIA discrete GPUs: works from 2GB VRAM (Qwen3-TTS) or 4GB (IndexTTS 2 / 2.5); the automatic mode still prefers CUDA, nothing to configure;
• AMD and Intel Arc discrete GPUs: accelerated through Vulkan, 4GB+ VRAM recommended—machines that could not install the originals at all;
• Macs with Apple silicon: macOS 15 or later; with 16GB of unified memory install the lite editions, 24GB+ can run the originals;
• Macs with Intel chips: macOS 15+ can install the lite editions, but only in CPU mode—clearly slower;
• Windows laptops with integrated graphics only: a CPU mode runs it—clearly slower, but enough to prove the workflow before any hardware decision.

What a “lite” edition actually changes ​

Nothing about the app was rebuilt: the interface, the workflow and the languages are identical to the original. Two things underneath are different.

The model is a GGUF-quantized version. Quantization is a tighter compression tier of the same model, not a smaller model (the explainer article covers the mechanics). In practice, VRAM and disk use drop sharply—IndexTTS 2 / 2.5 go from "NVIDIA with 8GB+" to any 4GB-class discrete GPU, and Qwen3-TTS Lite runs from 2GB of VRAM.

The acceleration backend moves from CUDA to Vulkan. CUDA is NVIDIA's private interface; Vulkan is an open, cross-platform graphics API—which is exactly why AMD and Intel Arc cards get hardware acceleration and Macs run natively.

Installing and running it ​

The flow is the same as every other app in the catalog, four steps.

Step 1: open the app's detail page in the PC client, check the description and system requirements, and click install. The three detail pages:

Qwen3-TTS Lite:

IndexTTS 2 Lite:

IndexTTS 2.5 Lite:

Step 2: pick an install location and start—models and environments download automatically, no command line:

Step 3: click launch and wait until the run window reports the service is ready.

Step 4: the browser then opens the working page—type text, pick a reference audio, hit synthesize. The three working pages:

Qwen3-TTS Lite:

IndexTTS 2 Lite:

IndexTTS 2.5 Lite:

Uninstalling is one click in the same client.

What survives, and what doesn’t ​

Emotion control in IndexTTS 2 / 2.5: the originals offer several ways to steer emotion; the lite editions keep the two workhorses—emotion-vector control (dial emotion type and intensity by parameter) and natural-language description control (type "questioning in anger" and it reads that way). Together with the voice's own default emotion, that covers everyday dubbing work. The one method that doesn't survive is emotion reference audio—pulling an emotion from a separate clip and layering it onto a cloned voice.

IndexTTS 2 / 2.5 handle very long text: a whole chapter or a long video script can be synthesized in one run without manual chunking—pair it with pause control and pronunciation fixes and the alignment work shrinks more than you'd expect.

Three ways to get a voice in Qwen3-TTS: beyond 3–5 second zero-shot cloning, there is Custom Voice—a set of ready-made voices you pick and use with no recording at all—and Voice Design, which creates a brand-new timbre from one sentence of description (“a gentle woman, bright tone”). The ~97 ms generation latency survives, so the near-realtime feel is intact.

The gap to expect: quantized audio sits slightly below the original on quality and similarity ceilings, most audible on long sentences and strong emotion. If you want the ceiling, install the original.

Original or lite: the GPU decides, in one line each ​

• NVIDIA, 8GB+ VRAM: install the IndexTTS 2 / 2.5 originals, nothing compromised; for Qwen3-TTS, the original is the pick from 4GB up.
• AMD or Intel Arc, 4GB+ VRAM: the lite editions are the only option with GPU acceleration.
• Apple silicon Mac: 16GB of unified memory or less—install the lite editions; 24GB+—the originals work too.
• Intel Mac: lite editions only, and CPU mode only.
• Windows laptop with integrated graphics only: CPU mode runs end to end—try it, then decide about hardware.