Thinking Machines has unveiled a new approach to human-AI interaction that could change how we communicate with machines. Their latest model, called TML-Interaction-Small, is a massive 276-billion-parameter mixture of experts system designed for real-time engagement. This development pushes the boundaries of current voice and video AI capabilities, making interactions more fluid and natural than ever before.

Breaking New Ground in Real-Time Multimodal Communication

The core innovation is a shift away from traditional turn-based systems toward continuous, simultaneous interaction. Unlike earlier models that process voice, video, and text separately, this new approach integrates all modalities into a seamless flow. The system can listen, speak, watch, and react instantly, without waiting for user turns or explicit cues.

One standout demo features streams of micro-interactions, each lasting about 200 milliseconds. Using encoder-free early fusion, the model processes images and audio together in less than 200ms, similar to Meta’s Chameleon. This enables a more natural, conversational experience where the AI can interrupt, proactivity respond, or even initiate dialogue based on ongoing inputs.

New Benchmarks and Capabilities Set by Thinking Machines

The team showcased the model outperforming existing systems on several benchmarks, including BigBench Audio, IFEval, and FD-bench. But beyond numbers, the real focus is on the model’s ability to handle complex tasks that require timing awareness and context understanding. For example, it can initiate speech at specific times or respond appropriately during code-switching situations.

Two new internal benchmarks, TimeSpeak and CueSpeak, highlight these strengths. TimeSpeak tests if the AI can start talking at user-specified moments, like reminding someone to breathe every few seconds. CueSpeak checks if the model can speak at the right moments, such as when a person switches languages. These tasks demonstrate the model’s deep understanding of timing and context in conversation.

Another impressive demo involved visual tracking and timed responses, like counting actions in videos or answering questions about ongoing scenes. These tests show the model’s ability to combine visual and auditory cues in a continuous, proactive manner—skills that are crucial for more natural AI assistants.

Implications for the Future of Human-AI Interaction

This development marks a shift from simple chatbots to more intelligent, multi-sensory systems. Experts say it could lead to AI that is more proactive and helpful, capable of understanding and reacting to ongoing situations without explicit commands. For example, an AI assistant could monitor your posture or activity and offer real-time feedback or assistance.

Thinking Machines also hinted at future plans involving background agents working alongside interactive models. These could enhance AI’s ability to handle complex, multi-tasking scenarios in real environments. The overall goal is creating AI that can think, watch, listen, and respond as seamlessly as a human.

This breakthrough raises the bar for what “realtime” means in multimodal AI systems. It emphasizes continuous awareness and interaction, rather than turn-based exchanges. As these models become more capable, they could find applications in areas like virtual assistants, online education, collaborative work, and more.

Overall, the new models from Thinking Machines showcase a bold step toward more natural, dynamic AI interactions. As development continues, expect AI systems to become more proactive, context-aware, and capable of engaging in human-like conversations across multiple channels.

Inspired by

• https://www.latent.space/p/ainews-thinking-machines-native-interaction

Sources

• kyutai.org
• smol.ai
• smol.ai
• thinkingmachines.ai
• openai.com