SourceVentureBeat•Stalecollected in 11m
Thinking Machines Previews Real-Time AI Interaction Models
#multimodal#full-duplex#low-latency#voice-videointeraction-modelsthinking-machinesmira-muratijohn-schulmanopenai
Real-time full-duplex AI models enable natural multimodal conversations—game-changer for interactive apps.
30-Second TL;DR
What Changed
Introduces 'interaction models' as native multimodal systems treating interactivity as core architecture.
Why It Matters
This advances AI towards natural human-like interactions, potentially transforming collaborative tools and real-time applications. For practitioners, it signals a new paradigm in model design prioritizing fluidity over batch processing.
What To Do Next
Visit Thinking Machines' blog to sign up for the limited research preview.
Who should care:Researchers & Academics
Key Points
- •Introduces 'interaction models' as native multimodal systems treating interactivity as core architecture.
- •Employs full-duplex multi-stream design processing 200ms input/output chunks simultaneously.
- •Uses encoder-free early fusion with lightweight embeddings for raw audio (dMel) and 40x40 image patches.
- •Achieves impressive benchmark gains and reduced latency over turn-based systems.
- •Limited research preview coming in months, wider release later this year.
Deep Insight
AI-generated analysis for this event — not the original article.
Enhanced Key Takeaways
- •Thinking Machines has secured $150M in seed funding led by Sequoia Capital and Andreessen Horowitz, specifically earmarked for building custom silicon infrastructure to support the high-throughput requirements of full-duplex interaction models.
- •The company is leveraging a proprietary 'Temporal State Compression' (TSC) layer that allows the model to maintain context across long-duration conversations without the exponential memory growth typical of standard transformer architectures.
- •Early developer documentation indicates that the API will support 'interruptibility' as a native feature, allowing the model to dynamically adjust its output stream in response to user audio input mid-sentence without requiring a full model reset.
Competitor Analysis
Architecture
- Thinking Machines (Interaction Model)
- Native Full-Duplex (200ms chunks)
- OpenAI (GPT-4o/Advanced Voice)
- Turn-based/Hybrid
- Google (Gemini Live)
- Turn-based/Hybrid
Latency
- Thinking Machines (Interaction Model)
- ~200ms (End-to-End)
- OpenAI (GPT-4o/Advanced Voice)
- ~320ms (Avg)
- Google (Gemini Live)
- ~400ms (Avg)
Input Processing
- Thinking Machines (Interaction Model)
- Encoder-free Early Fusion
- OpenAI (GPT-4o/Advanced Voice)
- Tokenized Multimodal
- Google (Gemini Live)
- Tokenized Multimodal
Pricing
- Thinking Machines (Interaction Model)
- TBD (Research Preview)
- OpenAI (GPT-4o/Advanced Voice)
- Tiered (Plus/Team/Ent)
- Google (Gemini Live)
- Tiered (Gemini Advanced)
| Feature | Thinking Machines (Interaction Model) | OpenAI (GPT-4o/Advanced Voice) | Google (Gemini Live) |
|---|---|---|---|
| Architecture | Native Full-Duplex (200ms chunks) | Turn-based/Hybrid | Turn-based/Hybrid |
| Latency | ~200ms (End-to-End) | ~320ms (Avg) | ~400ms (Avg) |
| Input Processing | Encoder-free Early Fusion | Tokenized Multimodal | Tokenized Multimodal |
| Pricing | TBD (Research Preview) | Tiered (Plus/Team/Ent) | Tiered (Gemini Advanced) |
Technical Deep Dive
- Encoder-free Early Fusion: Bypasses traditional modality-specific encoders (like CLIP or Whisper) by mapping raw dMel audio features and 40x40 image patches directly into the latent space of the primary transformer block.
- Full-Duplex Multi-stream Design: Utilizes a dual-head output architecture where the 'Audio Generation Head' and 'Reasoning Head' operate in parallel, allowing the model to generate speech while simultaneously processing incoming sensory data.
- dMel Embeddings: A custom-developed audio representation that preserves phase information, critical for maintaining natural prosody and emotional inflection in real-time speech synthesis.
- Memory Efficiency: Employs a sliding-window attention mechanism combined with a compressed state cache to keep inference costs within the 200ms latency budget.
Future ImplicationsAI analysis grounded in cited sources
Full-duplex AI will replace traditional turn-based voice assistants in enterprise customer service by Q4 2026.
The reduction in latency to near-human conversational speeds significantly improves user satisfaction metrics and task completion rates in high-volume support environments.
The shift toward encoder-free architectures will force a pivot in hardware acceleration requirements.
By removing the need for separate encoder passes, the computational bottleneck shifts entirely to the primary transformer block, favoring memory bandwidth over raw floating-point operations.
Timeline
2025-09
Thinking Machines incorporated by Mira Murati and John Schulman.
2026-01
Company closes $150M seed funding round.
2026-04
Internal 'Project Echo' achieves stable 200ms full-duplex latency in controlled environments.
- 2025-09Thinking Machines incorporated by Mira Murati and John Schulman.
- 2026-01Company closes $150M seed funding round.
- 2026-04Internal 'Project Echo' achieves stable 200ms full-duplex latency in controlled environments.
Weekly AI Recap
Read this week's curated digest of top AI events →
AI-curated news aggregator. All content rights belong to original publishers.
Original source: VentureBeat ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
The weekly digest
One email a week. Unsubscribe anytime.