News Flash: Dessa - the Conversation Model, Kimi K3.1 Pro, & Apple's 20B On-Device Model?

1. Dessa: The First Native Speech-to-Speech Model Trained Without Text

The News Highlight:

Dessa represents a fundamental shift in conversational AI by bypassing the traditional text-based training pipeline entirely. Developed by Block, this full-duplex speech model was trained from scratch using only raw human conversation audio, allowing it to understand and replicate the nuances of sound that text cannot capture. Unlike current voice assistants that convert voice to text (ASR), process it, and then convert it back to speech (TTS), Dessa operates natively in the audio domain. This enables the model to exhibit human-like behaviors such as laughing mid-sentence, gasping in surprise, and adjusting its volume or intonation based on context, all while maintaining a real-time, fluid conversation.

• Training Methodology: Trained from scratch on raw human speech sounds with zero text input or pre-trained text models.
• Full-Duplex Architecture: Supports synchronized speaker tracks, allowing for natural interruptions and real-time interaction without the "turn-taking" lag of traditional models.
• Emergent Behaviors: Demonstrates the ability to finish a user's sentences, follow verbal instructions (e.g., "speak up"), and express complex emotions like empathy and gratitude.
• Knowledge Retrieval: Despite having no text-based training, the model successfully retrieves facts and context learned purely from auditory patterns.

DO-AI Analysis:

From a first-principles architectural perspective, Dessa solves the "semantic loss" problem inherent in the ASR-LLM-TTS stack. Traditional pipelines discard prosody, rhythm, and emotional inflection during the text conversion phase, which are critical for high-stakes human interaction. By treating audio as a primary modality rather than a derivative of text, Dessa reduces latency and improves the "authenticity" of the interface. For engineering teams, the takeaway is clear: the next frontier of UX isn't just faster inference, but multimodal native architectures that eliminate the translation layers between human intent and machine response.

• Primary Source: Dessa - the Conversation Model trained only on speech

2. Moonshot AI Releases Kimi K3.1 Pro with 1M Context Window

The News Highlight:

Moonshot AI has announced the release of Kimi K3.1 Pro, a high-performance 3T-class model designed to push the boundaries of open-weight intelligence. The model is currently available via API, with a full open-weight release scheduled for November 4, 2026. Kimi K3.1 Pro distinguishes itself with a massive 1-million-token context window and significant optimizations in Reinforcement Learning (RL) across hundreds of downstream domains. The release also includes a "Highspeed" variant, which offers 2-3x faster inference, targeting responsive agentic workflows in coding, financial research, and 3D creation.

• Context Capacity: Features a 1-million-token context window, enabling the processing of massive codebases or long-form financial documents.
• Benchmark Performance: Scored 55 on Artificial Analysis, placing it ahead of several established proprietary models in the same class.
• RL Optimization: Utilizes a specialized RL framework and sandbox infrastructure to ensure stable, predictable returns across diverse task domains.
• Availability: Accessible via Kimi.ai and API today, with open weights arriving early November to support the developer ecosystem.

DO-AI Analysis:

The release of Kimi K3.1 Pro signals the continued commoditization of high-reasoning, long-context models in the open-weight category. The architectural focus on "joint RL across hundreds of domains" suggests a move away from general-purpose fine-tuning toward a more robust, agent-centric optimization. For developers, the 1M context window combined with a high-speed inference tier makes this a primary candidate for RAG-heavy (Retrieval-Augmented Generation) applications where low latency is non-negotiable. The trade-off to watch will be the memory overhead required to maintain such a large KV cache at scale.

• Primary Source: Kimi K3.1 Pro Tech Blog

3. Apple’s Core Advanced: 20B Parameters Running on 2GB RAM

The News Highlight:

Apple’s "Core Advanced" model, a 20-billion-parameter powerhouse integrated into macOS, is redefining how large models operate on edge devices. By utilizing a sophisticated weight-loading strategy, the model resides primarily in flash storage rather than active RAM. When a prompt is processed, the system dynamically activates only 1 to 4 billion parameters—a Mixture-of-Experts (MoE) style approach optimized for the Apple Neural Engine. Testing on M3 Max hardware reveals that while a standard 20B model would typically consume 15GB of RAM, Apple’s implementation requires only 2GB, significantly extending battery life and freeing up system resources for other tasks.

• Memory Efficiency: Uses only 2.0 GiB of memory for a 20B parameter model, compared to ~15GB for traditional open-model deployments.
• Dynamic Activation: Actively utilizes only 1–4 billion parameters per prompt, optimizing for energy and compute efficiency.
• Hardware Integration: Specifically tuned for the Apple Neural Engine, resulting in 1/3 the energy consumption of comparable open models.
• Performance Trade-offs: While highly efficient for dictation and Siri tasks, initial testing indicates weaker performance in complex mathematical reasoning compared to full-weight open models.

DO-AI Analysis:

Apple is demonstrating a masterclass in "Flash-to-Neural Engine" orchestration. By treating flash storage as an extension of the model's active memory and using selective parameter activation, they have solved the primary bottleneck of on-device AI: the RAM ceiling. This architecture proves that for edge deployment, parameter count is a vanity metric; what matters is the "active-to-resident" ratio. Engineering teams should look at this as a blueprint for mobile AI—prioritizing specialized hardware acceleration and intelligent weight management over raw model size to achieve production-grade performance on constrained devices.

• Primary Source: Exploring Apple's On-Device Model

Morning Executive Comparison Matrix