Dessa is Block’s research preview of a conversational speech model. Block says it learned from human conversation without text or pretrained models, and can listen while speaking. Its official page is useful evidence of the research direction, but does not yet provide the information needed for a general deployment guide.
For readers looking to try it, the immediate route is the official Dessa page. It presents a public waitlist and a separate early-access application for a small group of AI safety researchers. We did not submit either form or verify that an application receives access.
What Block actually specifies
Block describes audio input and audio output, simultaneous listening and speaking, and inference faster than real time. It says the model runs locally on a MacBook Pro with an M4 Max or better. The page does not state the model’s parameter count, required unified memory, downloadable checkpoint size or a reproducible inference benchmark.
The hardware statement therefore cannot be turned into a minimum RAM recommendation. A chip family alone does not identify the complete machine configuration, runtime or workload. Before buying hardware for this preview, obtain the actual deployment package and its supported configurations.
Block also says it trained Dessa using one NVIDIA GB200 NVL72 system. That training claim does not establish the resources needed for local inference or the cost of serving a concurrent group of users. Those are separate measurements.
Full duplex changes what an evaluation must capture
For a conversation model that listens and speaks at the same time, a transcript is an incomplete evaluation artifact. A response can contain suitable words but arrive too late, speak over a correction or fail to stop when the user interrupts.
An independent evaluation should retain synchronized input and output audio, with timings for interruptions and response starts. It should distinguish a short acknowledgment from an answer that prematurely cuts off the user. It should also record the model’s behavior when the user changes their mind halfway through a sentence.
Consider a proposed synthetic test: the user starts asking for a meeting on Tuesday, then corrects it to Thursday while the assistant has begun replying. The successful outcome is not simply a natural-sounding acknowledgment. The final confirmed date must be Thursday, and any application action must use the corrected date. This is a test design, not a claim about Dessa’s current capability.
Another useful test holds the words constant while changing delivery: a calm request, an urgent request and a sarcastic one. Reviewers can then judge whether the response is appropriate without assuming that emotional-sounding speech proves reliable understanding of a person’s condition.
Demonstrations and deployment evidence answer different questions
Block says its published clips are live recordings and their subtitles were added afterward. The page describes the model as choosing what to say and how to deliver it. We have not independently verified the recordings or used them as benchmark measurements.
Selected recordings can show examples the developer considers worth publishing. A deployment decision needs the failure rate across a defined task set, including routine and difficult conversations. Retain unsuccessful sessions, specify the selection method and distinguish software failures from model behavior.
For an application that controls other software, also preserve the link between a spoken request, the interpreted action and the actual tool result. A convincing voice must not substitute for confirmation that the requested operation happened. A preview without a documented tool interface should not be described as a ready replacement for an existing assistant integration.
What would make it ready to evaluate locally?
The useful next evidence is concrete: a documented installation route, supported operating-system and memory configurations, runtime version, model distribution terms and an inspectable evaluation on real-time conversations. API documentation and pricing would answer a different question for teams considering hosted use.
As checked on October 25, the official material supports a research-preview article and an access-status explanation. It does not support a claim that any reader can download the model, run it on an arbitrary Mac or reproduce a published latency result. Recheck the official page when an actual package or serving interface becomes available.
Reporting note: Prepared with AI assistance from Block’s primary publication. EyesTech has not accessed or independently benchmarked Dessa. Evaluation examples are proposed procedures. The featured image is an original AI-generated editorial illustration.
