Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
Data Science
Simon Willison's Blog·20d ago·5 min read
Inkling: Our open-weights model
TL;DR
Mira Murati's Thinking Machines Lab has launched Inkling, an open-weights multimodal model with 975 billion parameters, designed to process text, images, audio, and video. The model is trained on a massive dataset of 45 trillion tokens, with a smaller version, Inkling-Small, in development.
✦ Why It Matters
Engineers can start experimenting with Inkling's open-weights model to develop innovative multimodal applications today.
Key Takeaways
- Experiment with Inkling for tailored AI solutions.
- Consider using the Tinker platform for fine-tuning models.
- Stay updated on Inkling-Small's release for smaller applications.
- Evaluate Inkling against other open-weight models for your needs.
Full Summary
Inkling is a Mixture-of-Experts transformer model featuring 975 billion total parameters, with 41 billion active during operation. It has been trained on a vast dataset of 45 trillion tokens, encompassing text, images, audio, and video, and is licensed under Apache-2.0.
The model is intended as a base for fine-tuning rather than a top-tier model, making it suitable for developers looking to customize AI solutions. Additionally, a smaller variant, Inkling-Small, is in testing with 276 billion parameters.
The model's training data includes both public domain content and third-party datasets, although details are sparse. Inkling's multimodal capabilities and efficient architecture position it as a strong contender against other open-weight models, particularly those emerging from China.