Thinking Machines launches smaller, efficient open-source AI model, Inkling-Small
Thinking Machines has released Inkling-Small, a new open-source AI model that nearly matches the performance of its larger predecessor, Inkling, while being significantly smaller. Inkling-Small is a 276-billion-parameter multimodal reasoning model, featuring a permissive Apache 2.0 license. Despite Inkling having 975 billion parameters, Inkling-Small achieves a score of 40 on the Artificial Intelligence Index, just one point below Inkling's score of 41. This new model utilizes 12 billion active parameters per token, a substantial reduction from Inkling's 41 billion, yet maintains strong capabilities in coding, reasoning, and multimodal tasks. It accepts text, image, and audio inputs and can process a context window of up to one million tokens. The model's smaller footprint reduces compute requirements and inference costs, making it more accessible for enterprises with some GPU resources, though it still requires significant hardware. Thinking Machines has made the model's weights available on Hugging Face and offers fine-tuning via its Tinker API. To encourage adoption, the company is offering a limited-time 50% discount on API pricing. While Inkling-Small excels in areas like coding and multimodal workflows, it shows weaker performance in factual knowledge compared to its larger counterpart, necessitating verification for high-stakes applications. The model's architecture is a sparse Mixture-of-Experts design, activating only a fraction of its parameters for each token, enabling efficient processing. This release highlights Thinking Machines' rapid development cycle and commitment to open-source AI.
The introduction of Inkling-Small by Thinking Machines demonstrates a significant trend in AI development: optimizing performance and efficiency through architectural innovation rather than sheer parameter count. By employing a sparse Mixture-of-Experts approach, the company has managed to deliver a model that approaches the capabilities of a much larger predecessor with substantially reduced computational demands. This strategy addresses critical industry challenges related to deployment costs, accessibility, and environmental impact. The choice of an Apache 2.0 license further signals a commitment to fostering broad adoption and commercial use, contrasting with more restrictive custom licenses emerging elsewhere. While the model's reduced size makes it more practical for enterprise deployment, its continued reliance on significant GPU resources underscores that truly democratized, on-device AI remains a future milestone. The trade-off between broad performance and deep factual recall highlights the ongoing need for specialized models and robust verification systems in AI applications, especially in sensitive domains.
AI-generated to prompt reflection — not editorial opinion, not advice, not a statement of fact. How this works.