Skip to content
← Models
Open weightsKittenML/10 trustMetrics-checked
KittenTTS logo

KittenTTS

15M-parameter TTS in 25MB — realistic voices on any CPU, no GPU required.

Visit official page →

About

KittenTTS Nano is a 15-million-parameter text-to-speech model that fits in 25MB and runs entirely on CPU — no GPU needed. Distilled from Kokoro/StyleTTS2-style architectures, it delivers surprisingly natural speech for its size, with 8 expressive voices and adjustable speaking rate.

Apache 2.0 licensed. The use case is edge and resource-constrained deployment: mobile apps, Raspberry Pi-class devices, offline assistants, and anywhere a 3B-parameter model is overkill. Quality won't match large models, but the size-to-quality ratio is exceptional. Ships as ONNX (INT8 quantized) with CoreML conversions available from the community. Currently in developer preview — expect rough edges.

Specs

Context window

N/A (text-to-speech model)

Pricing

Free — open weights (Apache 2.0), runs on CPU.

Free quota

Free — download and run anywhere, no GPU needed (Apache 2.0).

API access

Self-hosted (kittentts PyPI wheel; ONNX; Hugging Face checkpoints; community CoreML).

Weights

Open weights

Released

June 1, 2025

Benchmarks

No standardized benchmarks. Community reception: widely praised size-to-quality ratio for CPU-only TTS; independent comparisons are informal (YouTube/blog demos). Currently in developer preview.

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

ttson-deviceedgevoice-assistantsaccessibility

Trust score — 7/10

/10 trustMetrics-checked

Scored 7/10 (metrics tier). Apache 2.0, a genuinely distinctive efficiency story (15M params, CPU-only, 25MB), and real community enthusiasm. Docked three points: (1) developer preview status — rough edges expected; (2) quality ceiling is inherently lower than large models; (3) small team, young project (2025), so longevity is unproven. The right choice when deployment constraints dominate quality.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
6.9/10
Maintenance · recency, releases, issue load
5.0/10
Openness · license + docs
8.8/10
Manual review · sanity read + red-flag check
7.0/10
  • · HF downloads 34,614 (peer-normalized)
  • · released 499 days ago
  • · manual review: exceptional size-to-quality for CPU-only; developer preview, rough edges; small young team; license: Apache-2.0

Sources