Inkling: Open-Weights Multimodal Model Announcement
With 975B parameters, Inkling provides a new foundation for customizable AI.
1Next-Gen Open Model
Inkling is an open-weights AI model developed with a Mixture-of-Experts architecture featuring 975B total parameters. This multimodal model combines text, image and audio processing capabilities, trained on 45 trillion tokens of data. We're sharing full weights to enable user customization for diverse needs.
2Scalable Model Family
Starting with Inkling-Small, we plan a family of models in different sizes. This lightweight version with 12B active parameters delivers strong performance with low cost and latency. Consistent training methodology across model sizes aims to provide solutions for various use cases.
3Multimodal Capabilities
The model natively processes text along with visual and auditory data. Its 1M token context window facilitates working with long-form content. Multimodal abilities enable flexible real-world applications, particularly promising for entertainment, education and creative industries.
4Customization Platform
Fine-tuning capability for Inkling is provided through the Tinker platform. Developers can customize the model with their own datasets. The Inkling Playground interface allows interactive testing of model behaviors. Even the model's self-fine-tuning capability has been demonstrated.
5General-Purpose Design
Rather than specializing narrowly, the model aims for balanced performance across multiple tasks. It demonstrates competence in coding, reasoning, visual understanding and more. This broad spectrum facilitates adaptation to different workflows. Benchmark results show competitive performance across evaluations.
6Agentic Features & Coding
Inkling shows strong capabilities in tool use and code generation. It ranks high among open-weights models on Design Arena's Agentic Web Dev leaderboard. Can create functional web applications in single attempts. Produces multi-page documents with cohesive designs.
7Collaborative Refinement
The model can work through extended feedback loops. Demonstrated by developing a multiplayer snake game through 40 refinement iterations. This capacity is critical for future human-AI collaboration. Enables continuous improvement of outputs.
8Controllable Thinking Effort
Users can adjust the cognitive effort the model expends on problem solving. This feature optimizes the cost-performance balance for specific use cases. Test-time scalability offers flexibility for different workloads. Enables efficient resource utilization.
9Safety & Reliability
Safety measures were prioritized during model training. Various evaluation methods ensure output accuracy and consistency. Being open-weights provides transparency and auditability advantages. Helps users understand model limitations and take appropriate precautions.
10Real-World Applications
Practical use cases demonstrated include creating travel journals and filling job application forms. Multimodal capabilities make it suitable for educational material preparation and content generation. Customizable architecture adapts well to industry-specific solutions.
11Developer-Friendly Tools
Inkling Playground in Tinker console provides an interface for developer interaction. Supported by API integrations and documentation. Offers necessary tools for understanding model behavior and developing customization strategies. Aims to build an ecosystem open to community contributions.
12Milestone in AI Evolution
Inkling represents a significant advance in open-weights multimodal models. Its customizable foundation model approach enables AI adaptation to broader use cases. Opens new possibilities for both researchers and industry applications. Carries potential to shape the future of human-AI collaboration.
Original article: Hacker News: Front Page — Inkling: Açık Ağırlıklı Çoklu Ortam Modeli Duyurusu