Kyutai's Pocket TTS shrinks to 1.9 million parameters — the day's second small-voice-model entry
From the AI digest for 2026-10-20
What happened: On October 20 (~10:40 AM ET), Yann LeCun relayed a Kyutai post: "Pocket TTS has been getting updates those last few days. While it was ~100M parameters, we release smaller versions, down to 1.9M: an 18.2M model with the quality of the original, a 6.9M model, and a 1.9M model that runs 103x faster than real time on one CPU core. All with voice cloning."
What's genuinely new: This is a debut, not a development — no Kyutai or Pocket-TTS thread existed on the record before this post. It lands the same day Mistral announced Voxtral 3 (section 5 above), making voice models the day's second small-model entry: Kyutai's axis is model size and CPU speed rather than a voice-agents API. The figures — original quality at 18.2M parameters, 103x real-time on one CPU core at 1.9M, voice cloning across all sizes — are the vendor's own claims, relayed secondhand through LeCun, with no benchmark table, audio sample, or availability detail in the window.
Watch: Independent listening tests of the three tiers; where the small models actually ship — products, harnesses, on-device tooling; and how the all-sizes voice-cloning capability interacts with the audio-provenance and watermarking watches already running in the record.
Links: LeCun RT of Kyutai