Back home

Tofu-1

x+web 0.94
Visit source

Intro

Tofu-1 is the full-duplex speech-to-speech model formally named by the voice-AI company Bland on 13 November 2026 (UTC). The official account @usebland posted at 18:03:38Z (02:03:38 on 14 November, Asia/Shanghai) “Introducing Tofu-1, Bland’s new speech-to-speech model,” and linked to the product page at https://www.bland.ai/speech-to-speech. The full company post said every sound in the demo video — the caller, the agent, and the background — was generated by Tofu-1; that high-stakes calls often happen in storms, at the side of a highway, or with kids crying in the background; and that the model was trained on the callers enterprises talk to every day, so agents can handle those conditions too. It invited readers to join the research preview.

Bland’s Isaiah Granet (@zaygranet, whose bio says phone operator @usebland) posted at 18:09:33Z (02:09:33 Asia/Shanghai) “Today we’re introducing Tofu-1,” adding that it listens while it speaks, reasons while it talks, and can tell a caller apart from everything else on a line, such as a TV or a crying baby.

The product page describes Tofu-1 as “Bland's new speech-to-speech model” and “Bland's full duplex speech-to-speech model,” and says it is the first release in Bland’s family of speech-to-speech models. The capabilities listed there include: understanding audio, not only words; telling speakers apart, recognizing who a phrase is meant for, and separating speech from a crying child, a television, highway noise, a hospital intercom, and keypad tones; generating a whole scene with several speakers, overlapping voices, and the room around them; listening and speaking at the same time so a caller can interrupt or correct it mid-sentence, as they would a person; and keeping the conversation going while it works on a request, without saying a task is done before it is. The page also describes a “TV Test”: a clear “Yes” from a television should not be taken as the caller’s confirmation.

On access, the product page says research-preview access to Tofu-1 is open to a limited number of enterprises — teams that handle high-stakes calls in banking, healthcare, insurance, and government — and that Bland will email applicants as access becomes available. Do not describe it as generally available to every Bland customer or to individual users, and do not describe it as the earlier Bland Speech v3 or Bland TTS v3.

Key points:

Explain

Why it matters

This is a new model formally named on 13 November 2026 by Bland’s official account and product page, not an unnamed feature update to an existing phone product. The company post at 18:03:38Z (02:03:38 on 14 November, Asia/Shanghai) falls inside this 00:32–04:32 scan window, and the previous 00:32 scan did not record it. The product page describes it as the first release in Bland’s family of speech-to-speech models.

This page only restates Bland’s official product page, the @usebland company post, and Isaiah Granet’s public description. Claims about full duplex operation, scene understanding and generation, reasoning while talking, and the research-preview scope come from those sources. This page does not independently evaluate the model.

Limits

The public record does not give parameter counts, training-data size, a standalone model card, formal API documentation, or a general-availability timetable. Do not describe Tofu-1 as already available to every customer or to individual users, and do not describe it as the same product as the earlier Bland Speech v3 or Bland TTS v3.

Usage

  1. Read the official product page at https://www.bland.ai/speech-to-speech , and cross-check the company post at https://x.com/usebland/status/2121302984802152941 and Isaiah Granet’s post at https://x.com/zaygranet/status/2121304474178429220 ;
  2. Per the official description, this is currently a research preview for a limited number of enterprises; apply on the product page, and Bland says it will email applicants as access becomes available — do not describe it as already open to all users;
  3. When citing full-duplex operation, scene separation and generation, reasoning while talking, or the TV Test, attribute them to the 13 November 2026 official Bland product page or official posts, and recheck the upstream source;
  4. Do not describe Tofu-1 as Bland Speech v3 or Bland TTS v3, and do not describe it as generally commercially available; this page only restates public materials — recheck the upstream sources before citing them externally.

Advertisement