Trending Now - Thinking Machines Lab Unveils Real-time AI Interaction Models

WHO:

Techmeme - Thinking Machines Lab, an AI startup founded by Mira Murati and John Schulman.

VentureBeat - Thinking Machines Lab, a company founded by former OpenAI chief technology officer Mira Murati and former OpenAI researcher John Schulman.

WHAT:

VentureBeat - Thinking Machines Lab's TML-Interaction-Small model is designed to handle simultaneous input and output processing, allowing it to engage in real-time interactions without waiting for user input to complete.

VentureBeat - The TML-Interaction-Small model achieved a turn-taking latency of 0.40 seconds, which is comparable to natural human conversation speeds.

VentureBeat - The model scored 82.1 on the IFEval (VoiceBench) benchmark, outperforming GPT-realtime-2.0 (81.7) and Gemini-3.1-flash-live (67.6).

VentureBeat - The model demonstrated a refusal rate of 99.0% on the Harmbench, indicating its ability to handle user requests effectively.

VentureBeat - Thinking Machines Lab's TML-Interaction-Small model achieved a turn-taking latency of 0.40 seconds, outperforming Gemini-3.1-flash-live (0.57 seconds) and GPT-realtime-2.0 (1.18 seconds).

VentureBeat - On FD-bench V1.5, TML-Interaction-Small scored 77.8, nearly doubling the scores of its primary competitors (GPT-realtime-2.0 minimal scored 46.8).

VentureBeat - The model's architecture allows it to process 200ms chunks of input and output simultaneously, enabling real-time interaction.

VentureBeat - The interaction models utilize a multi-stream, micro-turn design that processes 200ms chunks of input and output simultaneously.

VentureBeat - The model achieved a turn-taking latency of 0.40 seconds, outperforming Gemini-3.1-flash-live (0.57 seconds) and GPT-realtime-2.0 (1.18 seconds).

VentureBeat - On the FD-bench V1.5, TML-Interaction-Small scored 77.8, nearly doubling the scores of its competitors.

VentureBeat - On FD-bench V1.5, TML-Interaction-Small scored 77.8, nearly doubling the scores of its primary competitors (GPT-realtime-2.0 minimal scored 46.8).

VentureBeat - The model's architecture allows it to process 200ms chunks of input and output simultaneously, enabling real-time interaction.

VentureBeat - The TML-Interaction-Small model is a Mixture-of-Experts (MoE) model with 276 billion parameters and 12 billion active parameters.

VentureBeat - TML-Interaction-Small achieved a turn-taking latency of 0.40 seconds, outperforming Gemini-3.1-flash-live (0.57 seconds) and GPT-realtime-2.0 (1.18 seconds).

VentureBeat - On the FD-bench V1.5, TML-Interaction-Small scored 77.8, nearly doubling the scores of its primary competitors (GPT-realtime-2.0 minimal scored 46.8).

VentureBeat - The model's architecture allows it to process 200ms chunks of input and output simultaneously, enabling real-time interaction.

Techmeme - Thinking Machines Lab announced a research preview of interaction models that can think and respond in real time.

VentureBeat - The interaction models are designed to handle interaction natively, allowing for continuous user and AI interaction.

VentureBeat - The TML-Interaction-Small model has 276 billion parameters, with 12 billion active parameters.

VentureBeat - The model achieved a turn-taking latency of 0.40 seconds, outperforming Gemini-3.1-flash-live (0.57 seconds) and GPT-realtime-2.0 (1.18 seconds).

VentureBeat - On the FD-bench V1.5, TML-Interaction-Small scored 77.8, nearly doubling the scores of its competitors.

WHERE:

VentureBeat - The announcement and research preview are associated with Thinking Machines Lab, which is involved in AI research and development.

WHEN:

VentureBeat - A limited research preview of the interaction models will open in the coming months, with a wider release planned later in the year.

VentureBeat - A limited research preview will open in the coming months, with a wider release planned later this year.

WHY:

VentureBeat - The goal of the interaction models is to move away from "turn-based" interactivity to allow for more fluid and natural responses from AI.

VentureBeat - The models aim to solve the "collaboration bottleneck" that forces users to adapt their communication style to fit AI interfaces.

HOW:

VentureBeat - The system employs encoder-free early fusion, taking in raw audio signals and image patches through a lightweight embedding layer.

VentureBeat - The architecture includes an Interaction Model for dialog management and a Background Model for sustained reasoning and complex tasks.

VentureBeat - The model utilizes a dual system: an Interaction Model for dialog management and a Background Model for sustained reasoning and complex tasks.

VentureBeat - The system employs encoder-free early fusion, taking in raw audio signals and image patches through a lightweight embedding layer.

VentureBeat - The architecture includes an Interaction Model for dialog management and a Background Model for sustained reasoning and complex tasks.

VentureBeat - The model utilizes a dual system: an Interaction Model for dialog management and a Background Model for sustained reasoning and complex tasks.

VentureBeat - The system employs encoder-free early fusion, taking in raw audio signals and image patches through a lightweight embedding layer.

VentureBeat - The interaction models utilize a multi-stream, micro-turn design that processes 200ms chunks of input and output simultaneously.

VentureBeat - The architecture includes an Interaction Model for dialog management and a Background Model for sustained reasoning and complex tasks.

(UPDATED May 12, 2026 13:32 UTC)

Thinking Machines Lab has unveiled its TML-Interaction-Small model, which is designed for real-time interactions by processing simultaneous input and output without waiting for user input to finish. This innovative model achieved a turn-taking latency of just 0.40 seconds, closely mimicking the speed of natural human conversation, and scored 82.1 on the IFEval benchmark, surpassing competitors like GPT-realtime-2.0 and Gemini-3.1-flash-live. A limited research preview of these interaction models is set to open in the coming months, with a broader release planned for later this year.

(UPDATED May 12, 2026 11:37 UTC)

Thinking Machines Lab's new TML-Interaction-Small model has achieved a remarkable turn-taking latency of just 0.40 seconds, surpassing competitors like Gemini-3.1-flash-live and GPT-realtime-2.0. On the FD-bench V1.5, this innovative model scored 77.8, nearly doubling the performance of its closest rivals. Its advanced architecture allows for real-time interaction by processing 200-millisecond chunks of input and output simultaneously, utilizing a multi-stream, micro-turn design for efficient dialog management and complex reasoning tasks.

(UPDATED May 12, 2026 06:03 UTC)

Thinking Machines Lab's new TML-Interaction-Small model has achieved a remarkable turn-taking latency of just 0.40 seconds, surpassing its competitors, Gemini-3.1-flash-live at 0.57 seconds and GPT-realtime-2.0 at 1.18 seconds. On the FD-bench V1.5, this innovative model scored an impressive 77.8, nearly doubling the score of GPT-realtime-2.0, which only managed 46.8. The model's unique architecture allows it to process 200-millisecond chunks of input and output simultaneously, utilizing a dual system that includes an Interaction Model for dialog management and a Background Model for complex tasks.

(UPDATED May 12, 2026 04:14 UTC)

A groundbreaking development in artificial intelligence has emerged with the introduction of the TML-Interaction-Small model, which boasts an impressive 276 billion parameters and a remarkable turn-taking latency of just 0.40 seconds, outperforming its competitors Gemini-3.1-flash-live and GPT-realtime-2.0. Scoring 77.8 on the FD-bench V1.5, TML-Interaction-Small nearly doubles the performance of GPT-realtime-2.0, which scored only 46.8. This innovative model employs a dual system for dialog management and complex reasoning, utilizing encoder-free early fusion to process raw audio signals and image patches efficiently.

(UPDATED May 12, 2026 01:32 UTC)

Thinking Machines Lab, an AI startup founded by Mira Murati and John Schulman, has announced a research preview of innovative interaction models that can think and respond in real time, moving away from traditional turn-based interactivity. Their TML-Interaction-Small model boasts 276 billion parameters and achieved a remarkable turn-taking latency of just 0.40 seconds, outperforming competitors like Gemini-3.1-flash-live and GPT-realtime-2.0, while also scoring 77.8 on the FD-bench V1.5. A limited research preview is set to open in the coming months, with a wider release planned later this year, aiming to create more fluid and natural responses from AI.