Thinking Machines、初のAIモデル「Inkling」を公開

On July 15, Thinking Machines Lab announced Inkling, its first AI model, released under the Apache 2.0 license with full weights available on Hugging Face. The model is a Mixture-of-Experts Transformer architecture with 975 billion total parameters and 41 billion active parameters, supporting a maximum context window of 1 million tokens. Inkling was pretrained on 45 trillion tokens of multimodal data—text, image, audio, and video—enabling it to process and generate content across these modalities.

A distinctive feature of Inkling is its ability to adjust the computational effort devoted to reasoning, allowing users to control the trade-off between cost, inference latency, and output quality. This design prioritizes customization over raw performance; the company explicitly stated that "Inkling is not currently the strongest available model," instead positioning it as "a customization-friendly open-weight foundation" that combines multimodal capability, efficient reasoning, and ease of fine-tuning. During the announcement, the company demonstrated "self fine-tuning," in which Inkling itself writes, executes, and evaluates code for fine-tuning tasks.

Developers can begin customizing Inkling immediately through Tinker, Thinking Machines Lab's fine-tuning platform, which is being offered at a 50% discount during a limited promotional period. The Tinker console now includes Inkling Playground, an interface where developers can interact directly with the model to assess its behavior. To support deployment of fine-tuned versions, the company has partnered with inference providers TogetherAI and Fireworks.

Simultaneously, Thinking Machines Lab released a preview of Inkling-Small, a lightweight variant with 276 billion total parameters and 12 billion active parameters. According to the company, this smaller version achieves performance comparable to the full Inkling on many benchmarks, with full weights to be released after testing concludes. The dual-track rollout—a capable base model and a lighter alternative—underscores the company's focus on accessibility and adaptation across different resource constraints.

Thinking Machines Lab, founded by Mira Murati following her tenure as OpenAI's CTO, has entered the AI model space with a philosophy distinct from the pursuit of raw performance. Rather than claiming to offer the strongest available model, the company explicitly positions Inkling as a base designed for customization and adaptation. This strategy reflects a stated mission to "build AI that extends human will and judgment," emphasizing user agency over centralized capability.

The technical design of Inkling supports this positioning. Its Mixture-of-Experts architecture with tunable computational effort allows users to trade off speed, cost, and quality for their specific use cases—a feature particularly valuable for organizations that cannot deploy the largest models but need flexibility. The 45 trillion token pretraining corpus spanning text, image, audio, and video positions Inkling as a true multimodal foundation, though the company does not claim superiority on specific benchmarks.

The commercial model—open-source weights under Apache 2.0 license, coupled with a paid fine-tuning platform (Tinker) and partnerships with inference providers TogetherAI and Fireworks—suggests Thinking Machines intends to capture value through customization services and deployment rather than access restrictions. The immediate availability of a lightweight variant (Inkling-Small) and self-fine-tuning demos indicate the company is lowering barriers to adaptation, potentially making Inkling appealing to developers and organizations seeking control over their AI systems.

No comments yet. Be the first to share your thoughts!