AI is leaving the era "turn based" to chat?
Today, all of us who regularly use AI models for work or in our personal lives know that the basic mode of interaction over text, images, audio and video remains the same: a human user provides an input, between milliseconds and minutes (or in some cases, particularly difficult queries, hours and days) and the AI model provides the output.
But if AI is to truly take on the burden of jobs that require natural interactions, it must do more than provide such capabilities. "turn based" interactivity—as a result, it will need to respond more fluidly and naturally to human input, even while processing next human input, whether text or other format.
That seems like an argument to say the least Thinking machinesthe a well-funded AI startup It was founded last year by former OpenAI chief technology officer Mira Murati and former OpenAI researcher and co-founder John Shulman.
Today, the firm released a research preview of what it thinks is the case "interaction models, a new class of native multimodal systems that treat interactivity as first-class citizens of the model architecture rather than external software "trailer," making some impressive gains in third-party benchmarks and ultimately reducing latency.
However, the models are not yet available to the general public or even businesses, the company says in its statement announcement blog post: "We’ll open a limited research preview with a larger version later this year to gather feedback in the coming months."
‘Full duplex’ simultaneous I/O processing
At the heart of this announcement is a fundamental shift in how artificial intelligence perceives time and presence. Existing boundary models usually live on a tightrope; they wait for the user to complete input before they begin processing, and their perception freezes while they generate a response.
In their blog post, Thinking Machines researchers described the status quo as a constraint that forces people to do so "bend themselves" To AI interfaces, to express questions like email and gather ideas.
To solve this "cooperation bottleneck," Thinking Machines moved away from the standard alternate sign sequence.
Instead, they use a multi-threaded, micro-turnaround design that processes 200ms of input and output simultaneously.
This "full duplex" The architecture allows the model to listen, speak, and see in real-time, switching to the back channel when the user speaks or notices a visual cue, such as when the user makes a typo in a piece of code or a friend enters a video frame. Technically, the model uses encoder-free early coupling.
Instead of relying on massive stand-alone encoders like Whisper for audio, the system takes raw audio signals such as dMel and image patches (40×40) through a lightweight embedding layer, co-training all components from scratch within a transformer.
Binary model system
Provides an overview of the study TML-Contact-Smalla 276 billion parameters Expert Blend (TF) A model with 12 billion active parameters. Because real-time interaction often requires instantaneous response times that conflict with deep reasoning, the company created a two-part system:
-
Interaction model: Stays in constant communication with the user, managing dialogue management, presence and immediate follow-up.
-
Background Model: Continuous reasoning, web browsing, or an asynchronous agent handling complex tool calls flow results into an interaction model where the conversation will naturally be addressed.
This setup allows the AI to perform tasks like live translation or create a UI diagram while continuing to listen to user feedback—a capability demonstrated in the announcement video, where the model simultaneously generates a bar chart and provides typical human reaction times for various signals.
Impressive performance on key benchmarks against fast interaction models from other leading AI labs
A laboratory was used to prove the effectiveness of this approach FD-benchis a benchmark specifically designed to measure interaction quality rather than just raw intelligence. The results show that TML-Interaction-Small significantly outperforms existing real-time systems:
-
Liability: This queue has achieved a receive delay 0.40 secondsCompared to 0.57s for gemini-3.1-flash-live and 1.18s for GPT-realtime-2.0 (minimal).
-
Quality of interaction: Scored on FD-bench V1.5 77.8nearly doubled the scores of its main competitors (GPT-realtime-2.0 minimum score 46.8).
-
Visual activity: in specialized tests such as RepCount-A (counting physical repetitions in the video) and ProactiveVideoQAThinking Machines’ model successfully dealt with the visual world, while other frontier models remained silent or gave wrong answers.
|
Metric |
TML-Contact-Small |
GPT-real-time-2.0 (min) |
Gemini-3.1-flash-live (min) |
|
Queue delay (s) |
0.40 |
1.18 |
0.57 |
|
Interaction Quality (Average) |
77.8 |
46.8 |
54.3 |
|
IFEval (VoiceBench) |
82.1 |
81.7 |
67.6 |
|
Harmbench (rejection %) |
99.0 |
99.5 |
99.0 |
A potentially huge boon for businesses once the models are available
If introduced to the enterprise sector, Thinking Machines’ interaction models will represent a fundamental shift in how businesses integrate AI into their operational workflows.
A local interaction model such as TML-Interaction-Small enables several enterprise capabilities that are currently impossible or too fragile with standard multimodal models:
The current enterprise requires AI "convert" must be completed before analyzing the data. In a manufacturing or laboratory setting, a local interaction model can monitor video feed and proactively intervene the moment it detects a security breach or protocol deviation without waiting for an employee to ask for feedback.
The model’s success in visual benchmarks such as RepCount-A (counting accurate repetitions) and ProactiveVideoQA (answering questions as visual evidence appears) suggests that it can serve as a real-time auditor for high-risk physical tasks.
In voice-based customer service, the basic friction is 1-2 seconds "processing" General delay in standard APIs of 2026. Thinking Machines’ model achieves a queue delay of 0.40 seconds, roughly the speed of a natural human conversation.
Because the synchronous speech is handled natively, the enterprise support bot can listen to the customer’s frustration, "rear channel" signs (eg "i see" or "hmm-hmm") without interrupting the user and offer live translation that feels more like a natural conversation than a series of separate posts.
Standard LLMs do not have an internal clock; they are "to know" time only if provided in a text request. Interaction models are time-aware, allowing them to handle time-sensitive processes "Remind to check the temperature every 4 minutes" or "Let me know if this process takes longer than the last one". This is important for industrial maintenance and pharmaceutical research where time is an important variable.
Information about thinking machines
This release marks the second major milestone for Thinking Machines Launch of Tinker in October 2025A managed API for fine-tuning language models that allows researchers and developers to control their data and training methods while Thinking Machines manages the infrastructure load of distributed training.
Tinker supports both small and large open weight models, including expert mixture models, and early users include groups at Princeton, Stanford, Berkeley and Redwood Research, the company said.
Launched in early 2025, Thinking Machines bills itself as an AI research and product company that strives to make advanced AI systems “more widely understood, customizable, and generally competent.”
In July 2025, Thinking Machines said it had raised about $2 billion at a time worth 12 billion dollars in a round led by Andreessen Horowitz with participation from Nvidia, Accel, ServiceNow, Cisco, AMD and Jane Street Narrative as the largest seed funding round in its history.
The Wall Street Journal In August 2025, rival tech boss Mark Zuckerberg approached Murati about buying Thinking Machines Lab, and after he declined, Meta reported that more than ten of the startup’s roughly 50 employees were being pursued.
In March and April 2026, the company was also recognized for its computing ambitions: one Nvidia partnership to accommodate at least one gigawatt of next-generation Vera Rubin systems Expanded Google Cloud connectivity to leverage Google’s AI Hypercomputing infrastructure Model research, reinforcement learning workloads, boundary model training and Tinker with Nvidia GB300 systems.
Until April 2026 Business Insider reported on this Meta hired seven founding members from Thinking Machines, including Mark Jen and Yinghai Lu, while another Thinking Machines researcher, Tianyi Zhang, also transferred to Meta. The same report said Joshua Gross, who helped build Thinking Machines’ flagship fine-tuning product Tinker, has joined Meta Superintelligence Labs, bringing the company’s departure to about 130 employees.
Thinking Machines wasn’t just losing people, however: it also hired Meta veteran Soumith Chintala, creator of PyTorch, as CTO and added other high-profile technical talent like Neal Wu. TechCrunch Weiyao Wang, an eight-year Meta veteran working on separate multimodal perception systems, joined Thinking Machines in April 2026, stressing that the flow of talent is not one-way.
Thinking Machines has previously stated its commitment to this "significant open source components" in their publications to strengthen the research community. It is unclear whether these new interaction models will be subject to the same ethical and release conditions.
But one thing is certain: by making interactivity native to the model, Thinking Machines believes that scaling the model will now make it both a smarter and more effective collaborator.
