Dessa vs Gemma 4 12B: Two Laptop Models, Only One You Can Download

The most useful comparison between Dessa and Gemma 4 12B is not a benchmark score. It is a file. Gemma 4 12B — Google DeepMind's 11.95-billion-parameter open-weights model, published as Gemma 4 12B Unified under Apache 2.0 on 3 June 2026 — is a checkpoint you can download this afternoon, run on a 16 GB laptop, and fine-tune without asking anyone. Dessa, which Block announced on 23 October 2026, is a full-duplex speech model that runs on a MacBook Pro M4 Max according to its maker, except that you cannot download it, call it, or inspect it. It is behind a waitlist with no weights, no API, no price, and no published parameter count.

Both are sold as local models. Only one of them is actually distributed. That distinction organises everything else: what you can build, what you can measure, and what "runs on a laptop" even means.

The short answer, and why it is short

If your goal is to ship something this quarter that processes text, images, audio, or video, Gemma 4 12B is the only one of the two that is a product you can use. If your goal is to study full-duplex speech — a model that listens while it talks, interjects, and gets interrupted without losing the thread — Dessa is the more interesting artifact, and you cannot have it yet.

That sounds like a dismissal. It is closer to a description of two different stages of maturity. Gemma 4 12B is a finished open release with a licence, a model card, and a benchmark table. Dessa is a research preview with a demo reel. Comparing them on capability is like comparing a shipped camera to a lab prototype: the prototype may have the better sensor, but only one of them takes a photograph you can keep.

Two meanings of "local"

Gemma 4 12B's local story is concrete. It is a dense decoder-only model with an encoder-free unified architecture — 48 layers, image patches projected directly, audio handled without a separate encoder. Google positions it for roughly 16 GB of unified memory, and the weights are on Hugging Face in BF16 with GGUF, MLX, and other community builds following. You can run it offline, on a plane, with no account.

Dessa's local story is a sentence on a page: inference runs faster than real time on a MacBook Pro M4 Max or better. There is no memory figure, no quantisation, no file size, and no download. The claim may be entirely true. It is also unfalsifiable from outside, because "runs locally" in a vendor post can mean "our researchers demoed it on a laptop" rather than "you will run it on yours." Until there is a weight file or an installer, the honest reading is the first one.

One hears, one converses

The architectural difference is sharper than the parameter gap. Gemma 4 12B is a multimodal reader: text, image, audio, and video go in; text comes out. Its audio input is capped at 30 seconds per clip, and its video input at roughly 60 seconds at one frame per second. Within those limits it can transcribe, translate, describe, answer questions about a recording, and extract structured data. It is, in other words, extremely good at turning a bounded piece of the world into text.

Dessa is a conversational stream: audio in, audio out, with no text stage at all. There is no transcript to inspect, no token boundary, and no separate synthesis step. Block's claim is that this preserves prosody, laughter, sighs, and timing that a speech-to-text-to-speech pipeline flattens. Its demos show the model laughing mid-sentence, matching volume, and finishing a speaker's forgotten film title.

The practical consequence is that the two models are not substitutable at the API level. You cannot send Dessa a PDF. You cannot ask Gemma 4 12B to be interrupted mid-answer, because it does not produce an audio stream to interrupt. A voice product built on Gemma needs at least two more components; a document product built on Dessa is not possible at all.

The spec sheet, with both sides in the same line

Figures below are vendor-reported unless marked otherwise. Gemma's come from Google's model card; Dessa's come from Block's announcement page, which publishes far less.

• Parameters — Gemma 4 12B 11.95B dense vs Dessa undisclosed.

• Weights — Gemma 4 12B Apache 2.0 download vs Dessa none published.

• Context — Gemma 4 12B 262,144 tokens vs Dessa not stated, and no token interface to measure it against.

• Input — Gemma 4 12B text, image, audio up to 30 seconds, video up to about 60 seconds vs Dessa a live audio stream.

• Output — Gemma 4 12B text vs Dessa audio.

• Interaction — Gemma 4 12B turn-based request and response vs Dessa full-duplex, with barge-in and backchannel.

• Hardware — Gemma 4 12B roughly 16 GB of unified memory vs Dessa a MacBook Pro M4 Max or better, per Block.

• Price — Gemma 4 12B free to download; hosted rates at third-party providers cluster around $0.10 input and $0.30 output per million tokens vs Dessa no price.

• Benchmarks — Gemma 4 12B a published vendor table vs Dessa none.

• Availability — Gemma 4 12B generally available since June 2026 vs Dessa waitlist and a small safety-researcher preview.

Where Gemma 4 12B wins by default

Start with the numbers Google published, all vendor-reported and unreproduced here: MMLU-Pro 77.2, GPQA Diamond 78.8, AIME 2026 77.5 without tools, LiveCodeBench v6 72.0, Codeforces Elo 1659, Tau2 69.0, MMMU-Pro 69.1, and HLE 5.2. The pattern is a mid-sized model that is strong for its class on reasoning and vision, weak on the hardest frontier questions, and competent at tool use. A 256K context window and native function calling make it a plausible backbone for an agent that reads a long document and then acts.

Artificial Analysis, which is independent of Google, estimates the reasoning variant at 14 on its Intelligence Index — marked as an estimate rather than a measured run, and well above the median of 8 for comparable open models. That is a useful corrective to the vendor table: a mid-sized model, rated mid-pack by a third party.

Its audio numbers are more modest and more relevant to this comparison. CoVoST comes in at 38.5 and FLEURS at 0.069 (lower is better), both vendor-reported, with the CoVoST figure excluding Chinese. Those are speech translation and recognition metrics, not conversational quality. Gemma can hear a 30-second clip and tell you what was said. It cannot hold a live call.

The licence is the part that decides most commercial questions. Apache 2.0 means no revenue threshold, no acceptable-use negotiation, and no vendor permission for a fine-tune. For a team that wants to own its inference path, that is worth more than a better demo.

Where Dessa's demos point, if they hold

The interesting claim is not that Dessa sounds human. It is that removing the text bottleneck removes a category of failure. A pipeline that transcribes first has to decide when you have finished speaking; it commits to a transcript before it can respond; it loses the sigh. A duplex model never makes that commitment. If Block's clips are representative, the result is a conversation that tolerates overlap — you can talk over it, and it can talk over you.

What is missing is any way to know whether the clips are representative. Block says nobody steers the model and that subtitles were added afterwards, which rules out the crudest staging. It does not publish a held-out set, a latency distribution, an interruption-recovery rate, or an intelligibility score under noise. The Verge found the same demos "a little awkward and uncanny," including a clip where the model calls its partner "Ronaldo." A curated reel can survive that; a product has to.

Building a voice loop around Gemma today

If you want a voice agent now, the architecture is speech recognition, then Gemma 4 12B, then speech synthesis. The model handles understanding, reasoning, and the text of the reply; the surrounding services handle audio. The latency budget is dominated by the two edges, not the middle: endpoint detection plus synthesis usually costs more wall-clock time than the language model's first token.

OrcaRouter is useful in the middle of that loop for a different reason. One API in front of 200+ models means the reasoning step can be swapped without rewriting the application, and automatic failover means a provider hiccup does not end the call. The 0% markup model — provider list price passed through — matters when a vendor cuts a price, because the cut is live on our side the same day. One honest caveat: we route Gemma 4 26B A4B and Gemma 4 31B, not the 12B, which is a download-and-host model. If you want the 12B behind an API, that is a third-party hosting decision, not ours.

Choosing by constraint

• You need to ship text or multimodal output now — Gemma 4 12B, without a serious alternative in this pair.

• You need offline inference on hardware you control — Gemma 4 12B, because the weights exist.

• You need a document, an image, or a 30-second recording understood — Gemma 4 12B.

• You need a live voice that can be interrupted — Dessa is the only one of the two aimed at that, and it is not available.

• You need emotional expressiveness in the output audio — Dessa's demos are the evidence, and they are vendor-selected.

• You need a licence you can put through procurement — Gemma 4 12B's Apache 2.0 is the only licence on the table.

What would make this a real comparison

Three things would change the answer. A Dessa API, so the two could be measured on the same conversations. A published parameter count and latency figure, so the laptop claim could be reproduced. Or an open-weight speech-native model from someone else, which would turn "Dessa vs Gemma" into a comparison of architectures rather than a comparison of availability.

Until then, the fair summary is that Gemma 4 12B is a capable local multimodal model with a text bottleneck it cannot remove, and Dessa is a promising attempt to remove that bottleneck with no public way to check the result. The first is a tool. The second is a hypothesis with excellent production values.