Craitech has released TontaubeV1.1, a multilingual text-to-speech model with a four-bit option for smaller local systems. The TontaubeV1.1-int4 model card describes a 1.8 GB model download, style controls and support for cloning a voice from reference audio. The project tagged its v1.1.0 inference release on November 12.
The smaller download is a useful starting point for local voice experiments, but it is not the total storage requirement. The release documentation lists additional upstream models and runtime dependencies that can be downloaded on first use.
What changed
The publisher describes English and German as the primary languages, with additional support for Spanish, French, Italian, Dutch and Portuguese. Five controls adjust speaking rate, pauses, pitch range, pitch level and articulation. The release also adds a VibeVoice adapter and a four-bit version for smaller GPUs and CPUs.
For the default four-bit backend, the model card reports roughly 4 GiB of GPU memory at load and a peak around 6.5 GiB on supported NVIDIA Ampere-or-newer hardware. The project also documents a Linux x86-64 CPU path requiring AVX2, FMA and F16C instructions and a C++ compiler. Its higher-throughput full-precision vLLM backend has a substantially higher stated requirement: at least 24 GiB of VRAM.
These are publisher-reported requirements, not measurements from AiBrainWorX. The reference setup is Linux; the developers say Windows users can use CUDA-enabled WSL2, while native Windows is not supported. Compatibility and speed on a particular 12 GB GPU still need an actual test.
Weights and code have different terms
The model card identifies the Tontaube Community Model License 1.0 for the covered weights and says it is not an open-source license. The inference repository says its own source code is generally Apache 2.0 apart from identified third-party components. Check the model license, component notices and current terms for the intended use; do not assume the code license automatically applies to all weights.
The published license describes conditions for commercial use, attribution and offering speech-generation services. Those details can affect a product or hosted service even when a personal experiment is permitted.
Voice cloning and quality
Use only reference recordings you have permission to use. Voice cloning can enable impersonation, and an application should make clear when a voice is synthetic.
The project also publishes older evaluation figures for TontaubeV1 and explicitly warns that those results are not measurements of V1.1. AiBrainWorX has not installed this release, listened to its output or independently tested its memory use, speed or quality.
Original reporting from model card and release documentation; no independent installation or listening test.
