What was announced?
Block's AI research lab released Dessa, a speech model, as a research preview on October 23, 2026. Dessa is speech-to-speech: audio in, audio out, with no text layer in between.
The model is full-duplex, meaning it can listen while it talks. According to Block, it learned from real conversations when to cut in and when to respond with short cues like "yeah" or "got it."
What does speech-only training change?
Most voice systems chain three parts: speech recognition, a text model, then speech synthesis. Block's team argues that chain throws away emotion, sarcasm, laughter and sighs. Dessa processes those signals directly.
As far as Block has disclosed, the training data is human conversation audio only, with no text and no pretrained models. Training ran on a single NVIDIA GB200 NVL72 rack linking 72 Blackwell GPUs and 36 Grace CPUs. The company says the model learned language from hundreds of times fewer words than large language models need.
Limits
- Only demos exist so far; no benchmark scores have been published.
- Model weights, a technical paper, an API and pricing were not shared.
- Inference runs faster than real time on a MacBook Pro M4 Max or better. However, the project page and the lead researcher's social media post disagree on the hardware.
Comparison
Block compares Dessa with Kyutai's open full-duplex model Moshi, NVIDIA's PersonaPlex-7B-v1 and OpenAI's GPT-Live-1. Moshi starts from a text language model backbone and was trained on 1,016 H100 GPUs, while Dessa starts from speech directly. Block's figures are self-reported.
