Models#Text to speech#Model inference
Ollama adds local text-to-speech with Gradium's Phonon — English weights are Apache 2.0
Ollama's first speech model, Phonon, landed October 24: local TTS on Apple Silicon, OpenAI-compatible endpoint; English weights Apache 2.0, 40M noncommercial.
Ollama shipped its first speech model on October 24 with version 0.40.11: Phonon, a family of small text-to-speech models from Gradium, a Paris-and-Silicon-Valley voice AI lab founded by ex-DeepMind and Meta researchers. On an Apple Silicon Mac, upgrading Ollama and running ollama run phonon "Hello from Ollama." is enough to hear audio generated locally; applications get an OpenAI-compatible /v1/audio/speech endpoint. The licensing line was drawn on day one — the default phonon tag (a 31M-parameter English model) is Apache 2.0 and safe for products, while every 40M-parameter build (French, German, Spanish, Portuguese, and the five-language multilingual checkpoint) is CC BY-NC 4.0, noncommercial only, per the Ollama blog and library pages.
Key points
- What shipped: Phonon from Gradium is Ollama’s first speech capability, live since v0.40.11 (published 2026-10-24T00:30:13Z per the GitHub release); macOS only for now, with Windows and Linux “coming soon” by the official wording.
- The commercial split: the 31M English weights are Apache 2.0; all 40M builds are CC BY-NC 4.0. There is no licensed commercial path to non-English Phonon through Ollama today.
- Integration: one command speaks from the CLI; apps switch an OpenAI SDK’s base URL to port 11434 and stream 24 kHz mono WAV or raw PCM.
- Prior context: Gradium open-sourced the same weights on October 20; Ollama’s step is packaging — turning “downloadable” into “runs with one command.”
Background
Phonon did not appear for Ollama’s benefit. Gradium published the full suite on October 20 in a post titled “Open-Sourcing Phonon: A Suite of Frontier On-Device TTS Models”: 31M and 40M tiers, English under Apache 2.0, everything else CC BY-NC, weights on Hugging Face and inference code at gradium-ai/xn-ptts on GitHub. The launch framing was on-device first — 7 ms to first audio on a MacBook Air M5’s CPU, 38 ms and 3.7× real time on a Raspberry Pi 5, no GPU required by the company’s numbers. Gradium itself is a Kyutai-lineage startup; its July 2026 seed extension reached $100 million with NVIDIA joining the investor list.
Ollama’s contribution is runtime integration: the models joined its library (14 tags, 16,100 downloads within two days of the page going live), MLX acceleration on Apple Silicon, and speech wired into both the ollama run CLI and the HTTP API. For the local-model ecosystem this closes the loop on audio output — Recursive’s coding model Flint landed in the same library a few days earlier for text, and Phonon repeats that move for speech. What ollama run <model> supports is quietly becoming the de facto checklist of what runs locally.
The facts
Every performance figure below is Ollama’s own benchmark, unreplicated by third parties as of October 27: on an M5 Max MacBook Pro, over 45 seconds of speech with the median of 10 runs, Ollama through MLX was 35% faster than the next-fastest engine, Gradium’s own ptts runtime; word error rate matched the Hugging Face Transformers reference implementation on 200 Seed-TTS English sentences transcribed with Whisper large-v3. A second set comes from Gradium’s launch data: the 31M model scored 1.70% WER on a 500-sentence hard-text set (phone numbers, email addresses, postal codes), against Kokoro’s 3.50%.
| Item | Status (checked 2026-10-27) |
|---|---|
| Shipped | 2026-10-24, with Ollama v0.40.11 |
| Platform | macOS on Apple Silicon; Windows and Linux promised, no dates |
| Default model | phonon (31M params, English, Apache 2.0, ~104 MB download) |
| 40M family | French, German, Spanish, Portuguese, multilingual — CC BY-NC 4.0, all under 148 MB |
| Voices | One default voice on the English build; more via Gradium’s Hugging Face voice catalog |
| API | OpenAI-compatible /v1/audio/speech on port 11434, streaming, 24 kHz mono WAV or PCM |
| Input limit | 4,096 characters per request; ollama run splits longer text at paragraph and sentence breaks |
| Speed | Vendor-claimed 35% faster than Gradium ptts on M5 Max (45 s of speech, median of 10) |
The licensing split
The most consequential line in this release is the license boundary. How2Shout’s (H2S Media) analysis puts it bluntly: “There is no commercial path to non-English Phonon through Ollama today.” The 31M tier speaks only English; the moment a product needs French, German, Spanish, or Portuguese, the only weights available are CC BY-NC, and shipping them commercially breaches the license. The same analysis warns that the NC suffix is easy to misread as permission rather than restriction — “the error that costs money later.” Coverage itself is narrow: five languages, no Chinese, no Japanese, and no Indian languages; Gradium’s FAQ answers the gap with custom models on request.
What others say
Digital Trends (October 26) framed the release against subscription-locked TTS: the good synthesizers mostly sit behind paywalls, and free local options are rare. The author listened to Ollama’s published samples and found the voices “really natural” — an impression of vendor-picked samples, not an independent test; the article’s performance and licensing details likewise trace back to the official blog. How2Shout flagged practical quirks worth knowing before you build: Markdown markers are skipped when spoken, while prices, symbols, email addresses and URLs go the other way and get spelled out character by character, and dates read better written out (“October 23” rather than “10/23”). At least six outlets — 9to5Mac, Notebookcheck, Technobezz, Mezha, Jordan News among them — picked the story up within two days, mostly restating the official post; Chinese tech media followed on October 27 (Sina Finance, ifeng among them), also relaying official numbers. Hacker News has no substantive thread yet (an Algolia search on October 27 returned no submissions).
How to try it
After upgrading to 0.40.11, three commands:
ollama pull phonon
ollama run phonon "Hello from Ollama."
ollama run phonon "..." > speech.wav
To read a chat model’s reply aloud, pipe it: ollama run gemma4 --hidethinking "..." | ollama run phonon — text streams across, so playback starts after the first sentence rather than waiting for the full reply. For apps, point an OpenAI SDK at http://localhost:11434/v1/audio/speech; ollama show phonon lists every voice. The disk cost is small: 104 MB downloaded, about 121 MB resident for the English build.
Our take
For developers building voice features on a Mac, this release pushes the cost of trying local TTS down to one command: no subscription, no text leaving the machine, works offline, and near-zero code changes if you already speak the OpenAI audio API. What deserves scrutiny is the language-license combination — an English product can start today on the 31M weights, while a non-English product must choose between a noncommercial 40M prototype and negotiating with Gradium for custom models or its API, with no sign of that boundary moving. Two things remain unverified: the 35% speedup is the vendor’s own measurement, still awaiting an independent rerun on real hardware; and the quality gap between the multilingual and English builds rests on Gradium’s own benchmark (2.11% WER for 40M multilingual against 1.70% for 31M English). For contrast on direction, Block’s speech-to-speech model Dessa, which we covered on October 24, skips the text pipeline entirely; and for the recognition half of a voice pipeline, Microsoft’s streaming transcription model MAI-Transcribe-2 leads the latency charts but stays cloud-side, sold through Microsoft Foundry and OpenRouter rather than running next to Phonon on your Mac.