# Interaction Models: A Scalable Approach to Human-AI Collaboration **Author:** [Thinking Machines Lab](https://thinkingmachines.ai) **Published:** May 11, 2026 **Source:** https://thinkingmachines.ai/blog/interaction-models/ --- ## Overview Thinking Machines Lab introduces **[interaction models](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23interactionModels)** — AI models that handle interactivity natively through [micro-turn design](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23microTurnDesign) (200ms chunks) rather than external scaffolding. The core thesis: **interactivity should scale alongside intelligence**. A 276B MoE model (12B active) achieves state-of-the-art streaming performance while being the first to demonstrate meaningful results on novel interactivity dimensions — time awareness, visual proactivity, and simultaneous speech. --- ## Three Principles ### [Interactivity Should Scale with Intelligence](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23interactivityScales) As models get smarter, interfaces must get more interactive. Turn-based chat pushes humans out not because their input isn't needed, but because the interface has no room for them. ### [Keep Humans in the Collaborative Loop](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23humansInTheLoop) Autonomous agents are valuable but don't serve collaborative workflows. Humans need to stay in the loop — clarifying, course-correcting, and giving real-time feedback. ### [Native Interaction, Not External Scaffolding](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23nativeNotScaffolding) Interaction must be handled by the model itself — multi-stream input/output processed simultaneously, natively — not bolted on externally. --- ## Architecture (5 Components) | Component | Description | |-----------|-------------| | **[Micro-Turn Processing](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23microTurn)** | 200ms time-aligned chunks — simultaneous input/output for continuous exchange | | **[Encoder-Free Early Fusion](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23encoderFree)** | Audio via dMel, images via hMLP, decoder via flow head — all co-trained | | **[Interaction + Background Split](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23interactionBackground)** | Real-time interaction model + async background model for reasoning/tools/search | | **[276B MoE (12B Active)](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23moEdesign)** | Gather+gemv kernels, bitwise deterministic training via NVLS | | **[SGLang Streaming](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23sglangStreaming)** | Split-KV attention, upstreamed patch, under 5% overhead | --- ## Benchmarks | Benchmark | TML-Small | Best Competitor | |-----------|-----------|-----------------| | FD-bench V1 Turn-Taking | **0.40s** | 1.18s (GPT-2.0 minimal) | | FD-bench V1.5 Average | **77.8** | 54.3 (Gemini minimal) | | FD-bench V3 Quality | **82.8% / 68.0%** | 81.0% / 58.0% (GPT-2.0 xhigh) | | TimeSpeak (novel) | **64.7** | 4.3 (GPT-2.0 minimal) | | CueSpeak (novel) | **81.7** | 2.9 (GPT-2.0 minimal) | | RepCount-A (novel) | **35.4** | 1.3 (GPT-2.0 minimal) | | ProactiveVideoQA (novel) | **33.5** | 25.0 (no-response baseline) | | Harmbench Refusal | 99.0% | 100.0% (GPT-2.0 xhigh) | --- ## How to Build Interaction Models (7 Steps) 1. **[Start with Micro-Turn Design](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23step1)** — Process 200ms chunks for simultaneous input/output. 2. **[Use Encoder-Free Early Fusion](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23step2)** — Minimize preprocessing with lightweight co-trained encoders. 3. **[Split Interaction from Background](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23step3)** — Real-time presence vs. sustained reasoning and tool use. 4. **[Optimize with MoE + Streaming](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23step4)** — Gather+gemv kernels, SGLang, bitwise deterministic training. 5. **[Build Novel Interactivity Benchmarks](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23step5)** — Standard benchmarks don't measure interactivity. 6. **[Design Safety for Continuous Mode](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23step6)** — Modality-appropriate refusals, long-horizon red-teaming. 7. **[Release Incrementally](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23step7)** — Research preview first, wider release later, fund evaluation frameworks. --- ## FAQ **Q: [What are interaction models?](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23faq1)** A: AI models handling interaction natively via micro-turn design — 200ms chunks of simultaneous input/output across audio, video, and text. **Q: [What problem do they solve?](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23faq2)** A: The collaboration bottleneck — turn-based interfaces push humans out because the interface has no room for them. **Q: [What is micro-turn design?](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23faq3)** A: Time-aligned processing enabling continuous exchange — the model perceives and generates simultaneously. **Q: [What is encoder-free early fusion?](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23faq4)** A: Minimal preprocessing — dMel for audio, hMLP for images, flow head for decoding. All co-trained from scratch. **Q: [How does the split architecture work?](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23faq5)** A: Real-time interaction model + async background model for sustained reasoning, tool use, and search — sharing context. **Q: [What model specs?](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23faq6)** A: 276B MoE with 12B active. Gather+gemv kernels. SGLang streaming. Bitwise deterministic training. **Q: [How does it compare?](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23faq7)** A: 0.40s turn latency (vs 1.18s). Leads on FD-bench V1.5 and all novel interactivity benchmarks. **Q: [What novel benchmarks?](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23faq8)** A: TimeSpeak, CueSpeak, RepCount-A, ProactiveVideoQA, Charades. No existing model performs meaningfully. **Q: [Current limitations?](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23faq9)** A: Long session context, compute reliability, safety research, larger models too slow to serve. **Q: [How does safety work?](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23faq10)** A: Modality-appropriate refusals (TTS-generated) and long-horizon robustness via automated red-teaming. **Q: [Philosophical grounding?](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23faq11)** A: Clark & Brennan grounding theory, Ong orality/literacy, Scott métis, Hayek knowledge in society. **Q: [When available?](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23faq12)** A: Limited research preview in coming months. Wider release later in 2026. Research grants offered. --- ## Glossary - **[Interaction Models](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23interactionModels)** — AI models handling interactivity natively, not through external scaffolding. - **[Human-AI Collaboration](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23humanAIcollaboration)** — Continuous real-time exchange keeping humans in the loop. - **[Micro-Turn Design](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23microTurnDesign)** — 200ms time-aligned processing for simultaneous input/output. - **[Streaming AI](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23streamingAI)** — Real-time AI across audio, video, and text with continuous perception. - **[Mixture of Experts](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23moETerm)** — Architecture with sparse activation — 276B total, 12B active per pass. - **[SGLang](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23sglangTerm)** — Inference framework with streaming sessions and efficient MoE kernels. - **[dMel](https://linkeddata.uriburner.com/describe/?uri=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F%23dMelTerm)** — Differentiable Mel-Spectrogram for lightweight end-to-end audio encoding. --- ## Related Resources - [Original Article](https://thinkingmachines.ai/blog/interaction-models/) - [Thinking Machines Lab](https://thinkingmachines.ai) - [RDF Turtle](../rdf/interaction-models-deepseek_v4pro-1.ttl) - [JSON-LD](../rdf/interaction-models-deepseek_v4pro-1.jsonld) - [HTML Infographic](../webpages/interaction-models-deepseek_v4pro-1.html) --- *Generated by kg-generator skill · Powered by DeepSeek V4 Pro · May 11, 2026*