Mira Murati's First AI Model Is Built on China's Blueprint... Wild
Introduction to Inkling: A New AI Model
Overview of Thinking Machines Lab
- Thinking Machines Lab, founded by Mira Murati, former CTO of OpenAI, has released its first model called Inkling.
- The startup raised a record $2 billion seed round at a $12 billion valuation before launching any products.
Features of Inkling
- Inkling is a mixture of experts transformer with 975 billion parameters; only 41 billion activate for typical prompts, enhancing speed and cost-effectiveness.
- It can handle up to 1 million tokens in context and was pre-trained on 45 trillion tokens across various media types (text, images, audio, video).
Capabilities and Accessibility
Functionality
- Inkling can transcribe speech, follow spoken instructions, answer questions about recordings, describe images, and analyze charts.
- Outputs are text-based but include code and structured data; it is fully open weight under Apache 2.0 license.
Availability
- Complete weights are available on Hugging Face for download and fine-tuning without licensing fees.
Performance Metrics
Benchmarking Results
- On Humanity's last exam (text-only), Inkling scored 29.7%, trailing behind competitors like GLM 5.2 (40.1%) and Claude Fable 5 (53.3%).
- Despite lower scores on benchmarks like SWE-Bench Verified (77.6%) and Terminal Bench (63.8%), it excels in breadth and efficiency.
Unique Design Choices
Efficiency Features
- Inkling features a controllable thinking effort dial allowing developers to adjust speed versus accuracy based on task requirements.
- It performs well in practical applications such as web development tasks compared to other models.
Innovative Self-Fine-Tuning
Demonstrated Capabilities
- In tests, Inkling created functional applications from prompts and refined projects through iterative feedback processes.
- Notably, it demonstrated self-fine-tuning capabilities by writing its own fine-tuning job using Tinker.
Advanced Reasoning Abilities
Research Task Execution
- When tasked with researching the state of OpenWeights models and making predictions while building an interactive dashboard simultaneously, it effectively hedged uncertain information.
Epistemic Awareness
- Trained to flag uncertainty rather than hallucinate confidently; this design choice enhances reliability in responses.
Safety Measures Implemented
Risk Management Strategies
- Extensive safety training focused on internal behavior specifications across modalities; external testers verified performance against dangerous capabilities.
Architectural Insights
Mixture of Experts Design
- The architecture follows DeepSeek V3 principles with multiple routed experts per token for enhanced processing efficiency.
Implications of Model Release
Strategic Positioning
Inkling’s release highlights the competitive landscape between American closed labs and Chinese open labs regarding model performance metrics.
Business Model Considerations
- Thinking Machines focuses on platform revenue through Tinker rather than charging per token usage for the model itself.
Future Developments
Upcoming Variants
- A smaller version of Inkling is also being previewed that matches or exceeds its larger counterpart in several benchmarks while maintaining lower active parameters.