KittenML releases Kitten-TTS 2 with Apache 2.0 code and Community-Licensed weights, plus vendor-stated CPU real-time inference
KittenML (Stellon Labs) announced Kitten TTS 2 around Oct 21, 2026, described by the vendor as a state of the art 1.7B parameter text to speech model with in c…
Published on MyPrivateClaw
Oct 22, 2026, 4:44 PM UTC
Coverage date
Oct 22, 2026
Last updated
Oct 22, 2026, 4:44 PM UTC
News summary
KittenML (Stellon Labs) announced Kitten TTS 2 around Oct 21, 2026, described by the vendor as a state of the art 1.7B parameter text to speech model. The vendor states in context voice cloning from 5 30 seconds with no fine tuning, multilingual expression control, and CPU only real time inference via a llama.cpp fork; expression control is labelled beta and described as steering delivery rather than guaranteeing it. The code and model weights are separately licensed, with full terms below. The model body uses exact ternary (1.58 bit) packing of its linear weights to 1.6 bits per weight, which the vendor reports as 3.6x smaller at 954 MiB versus 3.47 GB in bf16. Because the packing is exact, the two produce bit identical audio; a separate 4 bit embedding variant (emb4) is the only lossy packing. Licensing splits across two artifacts, per the vendor's stated terms. The code is Apache 2.0…