Dessa: Conversation Model trained only on speech

Block says Dessa listens and talks at the same time, and inference runs faster than real time. The company reports that training used a single NVIDIA GB200 NVL72 and learned language from hundreds of times fewer words than LLMs. Block has posted conversation recordings but no benchmark numbers, and there is no public API or weights yet. The team is a small group inside Block. It was founded by former members of Dessa, the Toronto deep learning startup, which Block acquired in 2020. Most voice systems work differently: text goes in, text gets processed by a large language model, and then the output is synthesized into speech. Commenters on Hacker News questioned the team size and estimated the hardware cost of the single rack at several million dollars, a figure no one has confirmed. For founders, the claim is that a speech-first model can skip the text step entirely, which would change how teams handle latency, emotion and turn-taking. Until Block publishes benchmarks, pricing or weights, the results remain vendor claims. Teams building voice products should watch for independent evaluations and for when the public waitlist opens.