Thinking Machines Lab has introduced TML-Interaction-Small, a multimodal model designed to listen, watch and respond simultaneously rather than waiting for discrete conversational turns. The company says the model will enter a closed research preview in the coming months and expects to make it more widely available later in 2026.
Capabilities and architecture
- Input/output: concurrent audio, video and text in; concurrent audio and text out.
- Architecture: a mixture-of-experts transformer with 276 billion parameters total and 12 billion parameters active per token. A separate background reasoning model supports heavier inference; its precise architecture has not been disclosed.
- Two-part design: the system pairs a fast, interaction-focused model with an asynchronous background model. The interaction model operates using 200-millisecond "micro-turns" — it interleaves short chunks of input processing and output generation rather than following traditional alternating turns. As a result it treats audio, video and text as parallel streams and removes the artificial boundary between the end of input and the start of output.
Inputs and internal representations
The interaction model consumes:
- discretized audio tokens,
- embeddings of 40×40 pixel image patches produced by a hierarchical multilayer perceptron,
- text embeddings.
It generates audio and text using a flow-matching decoder. Thinking Machines Lab describes this as "encoder-free early fusion" because the team trained the transformer, perceptron and decoder jointly from scratch instead of relying on large pretrained encoders commonly used in multimodal systems.
The interaction model delegates more costly tasks — reasoning, web browsing and tool calls — to the asynchronous background model. Both components share the same conversational context, and the interaction model weaves background results into the live interaction when appropriate.
Benchmark performance
According to Thinking Machines Lab's tests, TML-Interaction-Small leads other voice-first models on interactivity benchmarks but lags behind GPT‑Realtime‑2's strongest reasoning mode on intelligence-focused evaluations.
Selected results:
- FD-bench V1 (audio latency in conversational turns): TML-Interaction-Small responded in 0.40 seconds, compared with Gemini‑3.1‑flash‑live‑preview at minimal reasoning (0.57 s) and GPT‑Realtime‑2 at minimal reasoning (1.18 s).
- FD-bench V1.5 (handling interruptions, interjections and foreground vs. background speech): TML‑Interaction‑Small achieved an average quality of 77.8, substantially higher than GPT‑Realtime‑2 at xhigh reasoning (47.8) and Gemini‑3.1‑flash‑live‑preview at high reasoning (45.5).
- Audio MultiChallenge (multi-turn audio dialogue, reasoning and instruction following): TML‑Interaction‑Small reached a 43.4% APR (average pass rate), behind GPT‑Realtime‑2 at xhigh reasoning (48.5% APR) but ahead of Gemini‑3.1‑flash‑live‑preview at high reasoning (36.1% APR).
- BigBench Audio (audio reasoning): with its background model active, TML‑Interaction‑Small scored 96.5% accuracy, slightly below the 96.6% achieved by GPT‑Realtime‑2 at high reasoning and Gemini‑3.1‑flash‑live‑preview at high reasoning.
Availability and undisclosed details
Thinking Machines Lab plans a closed research preview in the coming months and a broader release later in 2026. The company has not disclosed its training data and methods, the knowledge cutoff, the context window size, pricing, or the exact architecture of the background model.
Context and competitors
TML‑Interaction‑Small appears about 15 months after Mira Murati founded Thinking Machines Lab and is expected to be the startup's first public model. The company previously released a fine-tuning API called Tinker in October.
This year several firms announced real-time, audiovisual conversational models that handle interruptions: OpenBMB open-sourced the 9‑billion-parameter MiniCPM-o 4.5 in February; in March Google launched Gemini 3.1 Flash Live and Alibaba launched Qwen3.5 Omni; and OpenAI released GPT‑Realtime‑2 in May.
Of the models whose sizes have been disclosed, TML‑Interaction‑Small is among the largest trained specifically for interactive performance (276 billion parameters versus 9 billion for MiniCPM‑o 4.5, the closest architecturally similar public competitor). Thinking Machines Lab says it has larger pretrained interaction models but cannot serve them at real-time speeds yet and plans to release them later this year.
Why this matters
Many multimodal systems still impose noticeable pauses before responding or fail to react appropriately to real-time cues. Models that can listen, see and respond concurrently enable use cases that turn-based systems struggle with, such as live coaching or surgical monitoring. By combining a micro-turn interaction model with an asynchronous reasoner, TML‑Interaction‑Small explores a different trade-off between immediacy and deep reasoning in real-time multimodal agents.
Note on design approaches
The joint-training foreground/background approach of TML‑Interaction‑Small differs from orchestration strategies used elsewhere (for example, Vocal Bridge), where a real-time model defers heavy queries to a separate reasoner via tool calls. Orchestration can pair any real-time front-end with any reasoner without retraining, but it remains fundamentally turn-based and is constrained by API latency; the TML approach seeks learned handoffs between interaction and reasoning components.
