On 2026-10-23 (16:53 UTC) the official @blocks account unveiled Dessa, and Jack Dorsey called it the most authentic full-duplex speech model ever built. The official page block.xyz/dessa says Dessa is a true speech-to-speech model trained only on recordings of human conversation, with no text and no pretrained models. It is full-duplex, so it listens and talks at the same time, cuts in and backchannels naturally, runs faster than real time, and runs locally on a MacBook Pro M4 Max or better. Block says it was trained on a single NVIDIA GB200 NVL72 and learned language from hundreds of times fewer words than LLMs. There is no public API or weights yet; a small group of AI safety researchers is getting early access.
Key Takeaways
- ✓Trained only on human conversation audio, from scratch, with no text and no pretrained models (official page)
- ✓Full-duplex speech-to-speech: listens while talking, cuts in and backchannels, and natively hears and expresses emotion, sarcasm and laughter
- ✓Inference runs faster than real time and runs locally on a MacBook Pro M4 Max or better
- ✓Block says training used a single NVIDIA GB200 NVL72 and needed hundreds of times fewer words than LLMs
- ✓No public API, weights or third-party benchmarks yet; early access is limited to a small group of AI safety researchers
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and the problem it targets
Most voice AI today is either a cascade of speech-to-text, an LLM and text-to-speech, or speech bolted onto a text model. Tone, pauses, emotion and sarcasm are lost when speech becomes text, and replies often wait until the user stops talking, so conversations feel stilted and slow. Block argues this is also a safety problem: increasingly capable AI still cannot hear distress in a user's voice.
Architecture and how it works
Dessa comes from a very small AI research lab inside Block, founded by former members of the Toronto deep learning startup Dessa and reporting directly to Jack Dorsey. Block describes it as a true audio-in, audio-out model whose knowledge comes entirely from recordings of human conversation, with no text and no pretrained models in training. It is full-duplex, listening and speaking at once, and learned from real conversations when to cut in and when to backchannel; inference runs faster than real time. Block says training used a single NVIDIA GB200 NVL72 and that Dessa learned language from hundreds of times fewer words than LLMs.
Benchmarks
Block has published no standard benchmarks and there is no third-party evaluation. The launch page instead offers live recordings of people talking with Dessa (subtitles added afterward), showing it laughing mid-sentence, finishing a forgotten movie title and reacting to a surprise with what sounds like teary gratitude. One demo has it talking with GPT-Live-1, and Block notes that model's developer is not affiliated with Block and has not reviewed or endorsed Dessa. For now, any judgment rests on these official samples.
Getting started
It is not usable yet: there is no public API, weights or download. Block says more recordings will be shared on X and invites AI safety researchers to apply for early access via the contact on the official page. For teams building voice agents, two things are worth watching: whether a speech-only model running on-device can beat cascaded pipelines on latency and emotional understanding, and how such models get safety-tested, since Block explicitly treats this as new risk territory and wants outside researchers to probe it first.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.