Dessa

Dessa is a full-duplex speech-to-speech model developed by Block, built to process and generate raw conversational audio without relying on an intermediate text pipeline. Unlike conventional voice interfaces that transcribe speech to text, run it through an LLM, and feed the output to a text-to-speech engine, Dessa handles audio end-to-end. Audio goes directly into the model and audio comes out. This architecture preserves vocal inflection, pauses, interruptions, and emotional shifts that are stripped away when voice is flattened into text tokens.

Trained completely from scratch on human speech using a single NVIDIA GB200 NVL72 rack, Dessa bypasses pretrained foundation weights entirely. Block developed the system through its specialized AI research team, positioning the model as a fundamentally different category of conversational system. While mainstream audio agents lean on massive text datasets to bootstrap knowledge, Dessa learns language, conversational timing, and factual context purely from auditory signals.

For practitioners looking to deploy conversational intelligence on local infrastructure, Dessa establishes an important precedent: native full-duplex voice execution on edge hardware. Block demonstrated that the model runs locally on Apple Silicon, specifically targeting an Apple MacBook Pro equipped with an M4 Max chip or better. At present, Block has not released open weights, licensing terms, or API pricing, limiting early access to selected AI safety researchers.

Block built Dessa as a dense architecture with undisclosed parameters. Because the engineering team bypassed the standard text-tokenization paradigm, the model operates directly across raw acoustic representations.

In standard cascaded voice systems, three decoupled engines introduce additive latency: Automatic Speech Recognition (ASR), Large Language Model (LLM) generation, and Text-to-Speech (TTS) synthesis. This setup creates noticeable turn-taking delays and destroys acoustic context. If a user speaks with sarcasm, hesitation, or panic, the ASR layer discards that metadata, passing sterile text to the LLM.

Dessa's dense audio-to-audio design handles acoustic perception and acoustic generation within a single model loop:

• Full-Duplex Processing: The model processes continuous incoming audio while simultaneously generating outgoing audio streams. This enables natural backchanneling (such as interjecting with "yeah" or "got it") and instant interruptions without clearing an inference pipeline.
• Text-Free Training: By eliminating text completely, Dessa was trained on hundreds of times fewer total "words" than modern frontier LLMs. The model learned grammar, pronunciation, vocabulary, and conversational cadences simultaneously from acoustic variance alone.
• Compute Footprint: Block trained Dessa on a single NVIDIA GB200 NVL72 liquid-cooled rack. This demonstrates extreme compute efficiency for an end-to-end generative audio foundation model.
• Context Handling: While official context length specifications are not specified, full-duplex audio requires low-latency circular memory buffers to track acoustic history alongside speaker turns during live conversations.

Dessa is designed exclusively for natural, real-time verbal communication. Because it learned language exclusively through sound, its core strengths center on authentic dialogue execution rather than standard LLM workloads like code generation or document analysis.

Traditional voice bots wait for prolonged silence before triggering an endpoint and initiating inference. Dessa eliminates this dead air. It cuts in naturally, laughs mid-sentence when a joke is made, finishes a speaker's sentence if they hesitate, and yields immediately when interrupted.

Because Dessa natively hears raw pitch, timbre, speed, and volume, it responds directly to human emotion. In demonstrations, the model matched the speaker's emotional state, exhibited empathetic vocal tones, expressed surprise with audible gasps, and modulated its volume when instructed to speak up.

• Hands-Free Interactive Hardware: Robotics, vehicle cockpit control, and wearable devices requiring fluid, eyes-free back-and-forth interactions without server round-trip latency.
• Empathetic Interface Prototyping: Medical triage, crisis evaluation, and customer experience simulations where recognizing vocal stress, pain, or frustration is critical for safety.
• Low-Latency Edge Agents: Voice-driven gaming companions and desktop assistants that operate alongside local applications without cloud dependency.

Running full-duplex generative audio on local workstations presents unique hardware demands. Instead of simply generating discrete text tokens, the system must decode high-fidelity audio frames continuously while running real-time acoustic feature extraction on the microphone input.

Block explicitly highlighted the Apple MacBook Pro with an M4 Max chip as the reference baseline for running Dessa locally. The M4 Max features unified memory bandwidth reaching up to 546 GB/s, making it particularly capable of streaming continuous multi-track audio models without encountering the PCIe bottlenecks that can stall split CPU-GPU systems.

For developers targeting Apple hardware, 64GB to 128GB of unified memory provides the necessary headroom to store Dessa's weights while maintaining the active streaming buffers required for full-duplex generation.

While Block has kept the exact parameter count under wraps, evaluating how to run undisclosed model on consumer GPU hardware requires looking at modern dense speech systems. Based on the M4 Max baseline, Dessa VRAM requirements likely settle between 24GB and 48GB depending on runtime precision.

• Minimum Configuration (Quantized): A single 24GB GPU, such as an NVIDIA GeForce RTX 3090 or RTX 4090. If 4-bit quantization becomes available, 24GB of VRAM will be the entry threshold for local execution.
• Recommended Configuration (FP16 / High Precision): Dual RTX 4090 cards (48GB combined VRAM) or an enterprise card like the NVIDIA RTX 6000 Ada Generation (48GB VRAM). The best GPU for Dessa on dedicated Linux workstations is a high-memory, high-bandwidth card capable of sustaining sub-50ms inference cycles.

Quantizing speech-to-speech models is far more delicate than quantizing standard LLMs. In pure text models, 4-bit precision (such as Q4_K_M) retains reasoning with minimal perplexity degradation. For acoustic models, aggressive weight clipping can introduce audio artifacts, robotic distortion, or loss of subtle vocal inflections.

• Expected Quantization Tradeoffs: A 6-bit or 8-bit quantization format (like Q8_0 or FP8) will likely represent the best quantization for Dessa, preserving emotional nuance while fitting into standard workstation memory pools.
• Dessa Performance and Throughput: Rather than evaluating Dessa tokens per second in the traditional text sense, performance is measured in real-time factor (RTF). Dessa operates faster than real time (RTF < 1.0), meaning it generates audio frames faster than they are played back, completely preventing stutter or buffer underruns during live dialogue.

If Block releases Dessa weights to the public, local frameworks like Ollama, llama.cpp, or dedicated ONNX/vLLM-Omni runtimes will need updates to support native audio streaming primitives alongside standard weight formats.

Understanding where Dessa fits into the ecosystem helps teams decide between end-to-end audio models and composite LLM pipelines.

The industry standard for voice AI combines separate tools: Whisper for transcription, an LLM like Llama 3 for text processing, and a service like ElevenLabs for speech synthesis.

• Latency: Cascaded pipelines rarely drop below 500ms to 1000ms total latency. Dessa executes full-duplex inference in real time, dramatically cutting response delay.
• Acoustic Fidelity: Cascaded setups lose emotional context at the ASR layer. Dessa preserves vocal inflection, irony, and conversational cues because text is never introduced into the loop.
• Knowledge Breadth: Cascaded LLMs possess vast encyclopedic and coding knowledge. Dessa's knowledge is limited to what it has gathered from conversational human speech, making it ill-suited for deep programmatic or technical reference tasks.

Dessa enters direct competition with frontier speech models like OpenAI's GPT-Live-1 and Kyutai's open-source Moshi.

• Training Roots: GPT-Live-1 is anchored in a massive multimodal text foundation model that was subsequently aligned for voice. Dessa was trained strictly on sound with zero text pretraining.
• Hardware Independence: While GPT-Live-1 relies on proprietary cloud infrastructure, Dessa demonstrates functional edge deployment on consumer-accessible hardware like the M4 Max.
• Openness: Moshi provides open weights and accessible code, whereas Dessa remains restricted to safety researchers without published weights or open licenses.

For developers tracking local AI model undisclosed parameters 2026 developments, Dessa represents an important structural transition: moving past text-bound reasoning toward native acoustic intelligence running directly on local machines.