TL;DR
- Announced 23 October 2026: no weights, no code, no API, no licence and no benchmark numbers
- GitHub dessa org:block returns 0 results; Hugging Face authors block, block-xyz and blockxyz are all empty
- Three open full-duplex models ship weights today: Moshi (CC-BY-4.0), PersonaPlex-7B (gated) and VoiceChat 11B (OpenMDW 1.1)
System Requirements
| RAM | PersonaPlex-7B: about 16 GB VRAM at bf16 |
| GPU | VoiceChat 11B: A100, H100 or RTX-6000 on Linux with vLLM |
| VRAM | Apple Silicon: community MLX 8-bit, 4-bit and 3-bit builds |
Block announced Dessa yesterday and called it "the most authentic full-duplex speech model ever built." The lab lead posted that it is "a full-duplex speech-to-speech model that you can run on consumer GPUs" and that this is "a GPT3 moment for speech." You cannot run it. There are no weights, no code, no API, no licence and no benchmark numbers, and the only evidence on offer is a page of recordings Block selected.
What Block announced
Dessa is a speech-to-speech model: audio goes in, audio comes out, with no text stage that you are told about. It is full duplex, so it listens and talks at the same time, which is what lets it backchannel and get interrupted. Block says it was trained from scratch "on nothing but human conversation," with "no text, no pretrained models," on a single NVIDIA GB200 NVL72 rack. The page claims it learned language from "hundreds of times fewer words" than a language model needs.
The lab sits inside Block and reports to Jack Dorsey. It is led by Ragavan Thurairatnam, who founded and sold the Toronto deep learning startup that the model is named after. Access is an application form for a small group of AI safety researchers plus a public waitlist, with "launch links coming soon."
What Block did not ship
We checked the four places a released model turns up. All four are empty.
| Where we looked | What we found |
|---|
GitHub, dessa org:block | total_count: 0 |
Hugging Face, authors block, block-xyz, blockxyz | all three return an empty list |
| Licence | not stated anywhere on the announcement |
| Benchmarks | none published, against any harness |
That last row is the one that should bother you. Full-Duplex-Bench has existed since early 2025, it is public, and every open model below reports against it. Skipping it is a choice, not an oversight.
Why this is strange coming from Block
Block runs 143 public repositories under a GitHub organisation literally named "Open Source at Block, Inc." Its agent framework goose carries 55,637 stars under Apache 2.0 and was pushed to today, after moving out of the Block organisation into its own home in March 2026. Block sent Jack Dorsey to FOSDEM in 2025 to talk about putting open source culture into a company's DNA.
So this is not a company that does not know how to publish weights. It is a company whose AI lab, reporting to the same person, chose not to.
Full duplex, and why your voice stack is a walkie-talkie
Half duplex means one side transmits while the other waits. Every assistant built as speech recognition, then a language model, then text-to-speech is half duplex by construction, because the recogniser has to decide you have stopped talking before anything downstream can start. That decision is why assistants talk over you, and why they sit through a pause you did not mean as a handover.
Full duplex means both audio streams run the whole time. An interruption is not an exception the system handles, it is just something that happens in a stream. This is a genuinely harder architecture, and Dessa is right that it is the interesting one.
The one Dessa claim the open field cannot answer
Dessa's distinguishing claim is not full duplex. It is "no text." Every open full-duplex model routes through text internally, and their own documentation says so.
| Model | Text in the loop? | Evidence |
|---|
| Moshi | Yes | Paper: it predicts "time-aligned text tokens as a prefix to audio tokens," the Inner Monologue |
| PersonaPlex-7B | Yes | Built on Moshi, inherits it; the role prompt is plain text |
| NemotronLabs VoiceChat 11B | Yes | Card: audio tokens go to the backbone, which "predicts text tokens," which drive the TTS decoder |
So Block is pointing at a real gap. Two caveats keep it in proportion. Training a speech model with no text is not new, since Facebook AI Research published generative spoken language modelling from raw audio in 2021, and the open field moved to text-assisted architectures because they worked better. And the claim that dropping text produces a better conversation is exactly the claim that a benchmark number would settle.
Three open models with weights on disk today
All three are downloadable right now. Figures were read from the Hugging Face API on the date below.
| Model | Params | Licence | Gated | Downloads, 30 days |
|---|
| kyutai/moshiko-pytorch-bf16 | 7B | CC-BY-4.0 | No | 181,378 |
| nvidia/personaplex-7b-v1 | 7B | NVIDIA Open Model License | Yes | 98,185 |
| nvidia/NVIDIA-NemotronLabs-VoiceChat-11B | 11B | OpenMDW 1.1 | No | 5,174 |
Moshi is the ancestor: Kyutai shipped it in September 2024 with a stated 160 ms theoretical latency. PersonaPlex is NVIDIA's fine-tune of Moshi that added role and voice control, which we covered in detail in our PersonaPlex-7B spotlight. VoiceChat 11B is the newest and the least known, released on 3 August 2026, built from a Fast Conformer encoder, a Nemotron Nano v2 backbone and an NVIDIA TTS decoder.
The numbers Dessa does not have
Both NVIDIA models publish Full-Duplex-Bench results. These are vendor-reported, from separate runs, so read them as two self-assessments rather than a head-to-head.
| Full-Duplex-Bench 1.0 | PersonaPlex-7B | VoiceChat 11B | Dessa |
|---|
| Pause handling, synthetic (lower better) | 0.358 | 0.153 | not published |
| Pause handling, Candor (lower better) | 0.431 | 0.255 | not published |
| Smooth turn taking (higher better) | 0.908 | 0.82 | not published |
| Turn-taking latency | 170 ms | 448 ms | not published |
| Interruption takeover (higher better) | 0.950 | 1.00 | not published |
| Interruption latency | 240 ms | 480 ms | not published |
| Interruption answer quality, GPT-4o judge | 4.29 | 4.33 | not published |
The shape is a trade, not a winner. PersonaPlex is roughly twice as fast to speak; VoiceChat interrupts every single time, stays quieter during pauses by a wide margin, and matches it on answer quality with a bigger model. If your product breaks when the agent talks over a thinking pause, the 0.153 against 0.358 is the number to care about.
VoiceChat is the first open full-duplex model with native tool calling, emitted on its own output channel rather than bolted on afterwards. On Full-Duplex-Bench v3 it selects the right tool 82.5% of the time. Then it fills the arguments correctly 42.2% of the time, and completes the call end to end 33% of the time.
Read those three numbers in order, because the gap between them is the actual state of the art. Knowing which tool to reach for is close to solved; calling it correctly while holding a live conversation is not.
The licence spread
The three licences get progressively more corporate as the models get newer. Moshi is CC-BY-4.0 and ungated, which is as permissive as this field gets. PersonaPlex is the NVIDIA Open Model License and gated, so you accept terms and hand over contact details before the files move, which is awkward for a reproducible build or an air-gapped install.
VoiceChat is OpenMDW 1.1 and, notably, not gated. That is the same licence NVIDIA used on Nemotron 3.5 Bolt E4B yesterday, which makes two OpenMDW releases from the same vendor in one week. It is becoming NVIDIA's default, and it is worth reading once rather than assuming it matches the last one.
About that MacBook
Block says Dessa "runs locally on a MacBook Pro M4 Max or better." The X post says consumer GPUs. Neither is checkable, because nobody outside the application list has the weights.
The open side has the opposite problem and solves it in public. VoiceChat's card lists only datacentre NVIDIA parts and Linux with vLLM, which sounds restrictive until you look at what the community did in the days after release: GGUF builds, and MLX conversions at 8-bit, 4-bit and 3-bit for Apple Silicon. PersonaPlex wants about 16 GB of VRAM at bf16 and has MLX, GGUF, ONNX and WebGPU ports. Moshi has shipped MLX 4-bit weights since 2024.
Running one tonight
VoiceChat is the one to try first, because it is ungated and it is the only one that calls tools. Download the checkpoint and run the offline example, which takes a WAV in and gives you a WAV back without wiring up a microphone or TLS:
hf download nvidia/NVIDIA-NemotronLabs-VoiceChat-11B --local-dir voicechat
python "$NEMO_DIR/examples/speechlm2/offline_voicechat_infer.py" \
--checkpoint voicechat \
--wav "$NEMO_DIR/examples/speechlm2/sample_audio/sample_general.wav" \
--output-dir out
The environment setup is the fiddly part: a conda environment on Python 3.12, PyTorch 2.10, and the NVIDIA Speech repository on the nemotron-labs-voicechat branch, with causal-conv1d and mamba-ssm installed without build isolation. Budget an evening for the install, not the inference.
Who should care
If you are shipping a voice product this quarter, Dessa is not an option and will not be one until Block publishes something. Pick from the three above and treat the Dessa page as a statement of where the field might go.
If you need tools in the loop, VoiceChat 11B is currently the only open full-duplex model that does it natively. Plan around a 33% end-to-end pass rate, which means a confirmation step, not blind execution.
If latency is the product, PersonaPlex still answers an interruption in 240 ms, and nothing open is faster at a usable answer quality. Moshi is quicker still and has far less to say.
If you are watching the architecture question, Dessa is worth following, because text-free training is the one direction the open field abandoned and has not seriously revisited. Just hold the claims loosely until there is a number attached to them.
Sources and further reading
Your under-ten-minutes move: open the Dessa page, listen to one recording, and then go and download a model that exists. If you only have ten minutes and a Mac, take an MLX build of VoiceChat or Moshi and run the offline example, which tells you more about whether duplex speech is ready for your product than any demo reel will.
Not independently tested. Every Dessa claim above is Block's own, read from its announcement page and the lab lead's post on X, and none of it can be verified because the model is not available. Benchmark figures for PersonaPlex-7B and VoiceChat 11B are NVIDIA-reported, taken from the PersonaPlex paper and the VoiceChat model card respectively, and come from separate evaluation runs. Licence, gating status, download counts, release dates, repository stars and the absence of any Dessa code or weights were checked directly against the GitHub and Hugging Face APIs on the date below.
Date checked: 2026-10-24
