Block trained a voice model on one rack of hardware. The model never read a word of text. Block says it runs on a laptop.
The model is called Dessa, and Block announced it on October 23, 2026 01. Block says Dessa learned from human conversation alone, with no text and no pretrained model underneath. It is full-duplex. That means it listens and talks at the same time, like two people on a phone call.
Every clip anyone has heard so far came from Block. Block has released no public weights, no API, no price and no third-party benchmarks. So Dessa is a strong signal about where voice AI is heading. It is a weak signal about how good this particular model is today.
The stakes are plain money. A voice call that runs on the user's own machine costs the builder nothing per minute. Anyone paying a cloud provider for every spoken turn has a reason to care.
Speech carries what transcripts drop
Most voice agents run a relay. Speech gets turned into text, then a text model writes the reply. Block's Dessa page names the flaw bluntly: current voice AI "turns your voice into text and throws away an incredible amount of information text alone can't carry." A sigh and a half-second pause both vanish at the first handoff.
What each cent per minute costs a 1,000-user voice agent
Dessa is a bet on density. The idea is that a minute of real conversation holds more useful signal for talking than a page of text does. Block's claims fall into four parts.
On data, Block says Dessa used "hundreds of times fewer words" than text-based language models, all of it human speech. Block has not published how many hours of audio it used, or how many speakers it recorded.
On compute, Block says one NVIDIA GB200 NVL72 did the training. NVIDIA sells that system as a single rack holding 72 09 Blackwell GPUs and 36 Grace CPUs. One rack is still a data-center purchase, not a hobby budget. Block has not said how long the rack ran.
On inference, Block says Dessa runs faster than real time on a MacBook Pro with an M4 Max chip or better. Faster than real time means the model produces a second of audio in under a second. Block has not published latency in milliseconds or memory use.
On interaction, Dessa is full-duplex and it backchannels. Backchanneling is the little "yeah" or "got it" a listener drops in while you talk. Block's page says Dessa "laughs the moment something's funny, mid-sentence, the way people do."
Inside a voice stack without text
Picture the old relay as a phone call routed through a back office. A stenographer types what you say. A writer reads the transcript and drafts a reply. A voice actor reads that reply into the phone.
Each person in that office waits for the one before to finish, and nobody on the chain can type a sigh. Dessa collapses the office into one listener who hears the sigh and answers in kind.
Builders should judge any voice stack on one test: does the person on the other end stay on the line? TheStreet's Mehab Qureshi wrote about using a voice agent to book a restaurant table. The owner knew it was AI. "Not probably. Immediately," Qureshi wrote.
That hang-up is the conversion metric that matters. A booking agent that gets spotted in two seconds loses the booking. The relay keeps failing that test, and Block is pitching Dessa as the fix.
Step one is the latency budget. In a relay, delay stacks up at each handoff, and the agent also has to guess when you have finished talking. A full-duplex model skips the guess because it never stops listening. Block has not published a time-to-first-audio figure, so nobody outside Block can measure this gain yet.
Step two is the bill. Cloud voice typically charges on every stage of the relay, call after call. A local model moves that cost onto hardware the user already bought. The catch is the hardware itself: "M4 Max or better" is a high bar, well above a budget laptop.
Step three is the failure list. With no transcript in the middle, there is no text log to search or audit. A speech-only model may mishear a dosage or an address and keep talking smoothly. Full duplex can also cut off people who pause longer, including people with speech impairments.
Builders noticed the economics first. On Hacker News, the launch thread drew 503 12 points and 134 comments. The top comment, from user toddmorey, singled out the line about running locally on an M4 Max and training on one GB200 NVL72.
We think timing is the product here. A small model that answers on the beat may beat a smarter model that answers late. That could hold for booking calls and language practice. It is less clear for support lines, where getting the account number right matters more than sounding human.
Cheap minutes, unproven model, slow hardware curve
Every cent per minute is a yearly line item.
At 1,000 26 users talking 10 minutes 25 a day, each cent of per-minute cloud pricing costs $36,000 a year. Local inference pushes that line toward zero for users whose hardware can carry the model.
The M4 Max bar may drop to mid-range machines.
Block's stated floor is an M4 Max MacBook Pro or better, well above a budget laptop. If chips keep improving at the recent Apple pace, that floor could reach mid-range machines by 2031.
Every clip heard since launch came from Block.
Block has released no public weights, no API, no price and no third-party benchmarks. With no transcript in the middle, a speech-only model may also mishear a dosage or an address and keep talking smoothly.
Where voice runs shifts by 2031
The story starts in 2020, when Block bought the original Dessa AI startup, according to Gate News. Six years later, Block's AI lab, led by Ragavan Thurairatnam, ships a speech model whose stated target is a laptop. Watch where inference lives. In one announcement, it moved from a cloud endpoint to a single high-end machine.
Say each user talks to your agent for 10 minutes 25 a day. That is 300 minutes a month per user. At 1,000 26 users, you pay for 300,000 minutes a month.
Every cent your provider charges per minute adds $3,000 27 a month to that bill. Over a year, each cent is $36,000 28. Plug in your real per-minute price and multiply.
Local inference pushes that line toward zero for every user whose hardware can carry the model. That sets up counterpositioning. A cloud provider that bills by the minute cannot give away local voice without cutting its own revenue. A startup with no per-minute revenue loses nothing by doing it.
Five years takes us from 2026 to 2031. If laptop chips keep improving at the annual pace of recent Apple releases, the "M4 Max or better" floor could reach mid-range machines inside that window. Local voice might then ship as a default feature in laptops and cars. Or it might stall, because broad scale may still win on facts and reasoning.
The risk is lopsided. Testing local voice costs a builder a weekend of work. Ignoring it could cost the whole margin if a competitor ships the same agent with no minute bill. A small, early bet looks right.
Record twenty interruption tests before Monday
Dessa has no public weights yet, so nobody outside Block can run it today. You can still build the yardstick you will measure it with. The goal is one test set you can rerun on any voice model, cloud or local.
Step one: write 20 30 short exchanges in a doc. Make 5 where you interrupt the agent mid-answer and 5 where you say "yeah" while it talks. Make 5 that include a phone number or a spelled-out name. Make 5 with a TV playing in the background.
Step two: run all 20 30 against your current voice agent. Record both sides on one track with a second device in the room. Open the file in Audacity, a free audio editor, and read the gaps off the waveform.
Step three: log time to first audio for each clip. Time to first audio is the gap between your last word and the agent's first sound. Put every number in one spreadsheet column.
Step four: score each clip pass or fail on three questions. Did it stop when you cut in? Did it get the number right? Did it stay quiet when you went silent?
Step five: put your minute math next to the scores. Use the formula from above: users times daily minutes times 30 31 times your per-minute price. That cell is the cost a local model would have to beat.
Step six: join the waitlist at block.xyz/dessa. When access arrives, rerun the same 20 30 clips and compare columns. Expect plenty of clips to fail on your current stack. Those failures are your baseline, and the spreadsheet will show whether a local model earns its place.
Build a 20-clip voice test set you can rerun on any model
- Write the 20 30 exchanges. Draft 5 interruptions, 5 "yeah" backchannels, 5 with a phone number or spelled name, and 5 with a TV playing in the background.
- Record and time your current agent. Capture both sides on one track with a second device, open the file in Audacity, and log time to first audio for each clip in one spreadsheet column.
- Score and price the baseline. Mark each clip pass or fail on cut-in, number accuracy and silence, then add users times daily minutes times 30 31 times your per-minute price as the cost a local model must beat.
Measure your relay today so Dessa has a real yardstick when access opens.
Dessa points to voice agents that listen and talk at once on a single high-end laptop. Block has shown the direction and kept the evidence in-house. We think a small, early bet is the right move: a weekend of testing costs little, and a competitor with no minute bill could take the margin. Build the 20-clip 30 baseline, join the waitlist, and let the spreadsheet decide.