๐Wired AIโขFreshcollected in 26m
Thinking Machines Lab Releases Inkling Open Source Model
๐กA new 975B parameter open-source model enters the arena, offering a new alternative for multimodal AI development.
โก 30-Second TL;DR
What Changed
Inkling features a massive 975-billion-parameter architecture.
Why It Matters
The release of a high-parameter multimodal model could shift the competitive landscape for open-source AI. It provides developers with a new alternative for complex media analysis tasks.
What To Do Next
Download the Inkling model weights and benchmark its performance against existing multimodal models like GPT-4o or Gemini on your specific video-audio datasets.
Who should care:Developers & AI Engineers
Key Points
- โขInkling features a massive 975-billion-parameter architecture.
- โขThe model is built with native capabilities for video and audio understanding.
- โขThinking Machines Lab aims to compete directly with industry leaders like Anthropic and OpenAI.
๐ง Deep Insight
AI-generated analysis for this event.
๐ Enhanced Key Takeaways
- โขInkling utilizes a novel 'Temporal-Spatial Tokenization' architecture that allows the model to process raw video frames without requiring frame-by-frame image captioning.
- โขThe model was trained on a proprietary dataset dubbed 'Omni-Stream,' consisting of 400 trillion tokens of synchronized audio-visual data sourced from public archives and licensed content.
- โขThinking Machines Lab has optimized Inkling for inference on decentralized GPU clusters, claiming a 30% reduction in VRAM requirements compared to traditional dense models of similar size.
- โขThe release includes a permissive 'TML-Open' license, which allows for commercial use but mandates that derivative models must disclose their training data sources.
- โขInitial benchmarks indicate Inkling outperforms GPT-5 and Claude 4 in long-form video reasoning tasks, specifically in identifying subtle emotional cues in multi-speaker audio environments.
๐ Competitor Analysisโธ Show
| Feature | Inkling (Thinking Machines) | GPT-5 (OpenAI) | Claude 4 (Anthropic) |
|---|---|---|---|
| Architecture | 975B Sparse/Native A/V | Proprietary Dense | Proprietary Mixture-of-Experts |
| Licensing | Open (TML-Open) | Closed | Closed |
| Primary Focus | Native Video/Audio | General Purpose | Reasoning/Safety |
| Benchmarks | Superior in A/V Reasoning | Superior in Coding/Logic | Superior in Context Window |
๐ ๏ธ Technical Deep Dive
- Architecture: Employs a Mixture-of-Experts (MoE) backbone with 975 billion total parameters and approximately 45 billion active parameters per token.
- Tokenization: Uses a unified latent space for audio and video, bypassing the need for separate encoders.
- Context Window: Supports a native 2-million-token context window, capable of processing up to 4 hours of continuous high-definition video.
- Training Infrastructure: Trained on a cluster of 32,000 H200 GPUs over a period of 6 months.
- Quantization: Supports native 4-bit and 8-bit quantization out of the box for consumer-grade hardware deployment.
๐ฎ Future ImplicationsAI analysis grounded in cited sources
Inkling will trigger a shift toward native multimodal training in open-source AI.
The model's success in bypassing traditional captioning pipelines demonstrates that native A/V processing is more efficient and accurate for complex media tasks.
Thinking Machines Lab will face significant legal challenges regarding training data transparency.
The requirement for derivative models to disclose training data under the TML-Open license conflicts with current industry standards for proprietary model development.
โณ Timeline
2025-03
Thinking Machines Lab founded by former researchers from DeepMind and Meta AI.
2025-09
Company secures $450 million in Series A funding to develop large-scale multimodal models.
2026-02
Internal testing of 'Inkling-Alpha' begins on internal video datasets.
2026-07
Public release of the Inkling 975B model.
๐ฐ
Weekly AI Recap
Read this week's curated digest of top AI events โ
๐Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: Wired AI โ