Inkling: Our open-weights model

2026-07-16 · Source: Simon Willison's Weblog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Intermediate, short

Summary

Thinking Machines Lab released Inkling on July 16, 2026, their first open-weights multimodal model. Licensed under Apache-2.0, Inkling is a Mixture-of-Experts transformer featuring 975 billion total parameters with 41 billion active, trained on 45 trillion tokens encompassing text, images, audio, and video. While not positioned as a frontier model, it is designed as a robust base for fine-tuning, particularly through their Tinker training platform. The model card and training data documentation are notably brief, offering limited specifics on data sources beyond general mentions of public domain, open internet, and third-party content. Inkling is considered a competitive addition to the US open-weights ecosystem, alongside models like NVIDIA Nemotron and Gemma 4, demonstrating multimodal capabilities and efficient processing. An upcoming Inkling-Small, with 276 billion (12 billion active) parameters, is also planned.

Key takeaway

For AI Engineers evaluating open-weights models for custom applications, Inkling presents a viable Apache-2.0 licensed multimodal base model. You should consider its 975B total parameters and fine-tuning potential on the Tinker platform, especially if your project requires multimodal capabilities and efficient processing. Be aware that its training data documentation is minimal, which might impact your due diligence. Test its performance against your specific use cases, particularly for image generation and description tasks.

Key insights

Inkling is an Apache-2.0 licensed multimodal base model for fine-tuning, not a frontier model.

Principles

• Open-weights models can compete globally.
• Multimodal capabilities enhance base models.
• Data documentation often lacks transparency.

Method

The article demonstrates API interaction for multimodal models using `curl` commands to generate SVG images and then describe them via base64 encoded JPEGs.

In practice

• Use `curl` for API interaction.
• Test multimodal model image generation.
• Evaluate model's self-description accuracy.

Topics

• Open-weights Models
• Multimodal AI
• Mixture-of-Experts
• Fine-tuning
• Tinker Platform
• Apache-2.0 License

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Engineer, Machine Learning Engineer, AI Scientist

Related on AIssential

• not much happened today — Inkling's launch emphasizes an open-weight, customizable multimodal foundation model designed for practical application and ecosystem integration over pure benchmark leadership.
• Mistral Medium 3.5 - Mistral AI — Mistral Medium 3.5 is a multimodal, open-weight model optimized for agents and coding with extensive features.
• Mira Murati’s Thinking Machines drops Inkling, an open-weights model anyone can access — Open-weight multimodal models like Inkling offer customizable, cost-effective alternatives to proprietary AI systems.
• Complete Guide to Thinking Machines Inkling — Inkling is an open-weights multimodal MoE model prioritizing customization and domain-specific fine-tuning over benchmark supremacy.
• [AINews] Thinky's Inkling: 975B-A41B multimodal, new best American Apache 2.0 open model (with Inkling-Small, 276B-A12B) — Inkling is an Apache 2.0 multimodal MoE model, prioritizing customization and efficiency over raw benchmark supremacy.
• New Model: Inkling by Thinking Machine on Hugging Face — Multimodal models can achieve robust reasoning across image, text, and audio with simplified encoder architectures.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Simon Willison's Weblog.