Sinking Machines unveils 'Inkling-Small'... Achieves class-leading performance at one-quarter the size

Thinking Machines Lab (TML) has released its open-weight multimodal model, Inkling-Small, on Hugging Face.

Though roughly one-quarter the size of Inkling, it preserves most reasoning, coding, and multimodal performance—and beats it on some benchmarks.

It adds MoE architecture, variable reasoning, up to a 1 million-token context window, and support for image and audio inputs.

The model was further improved with reinforcement learning and is released under the Apache 2.0 license for modification, retraining, and commercial use.

It offers a practical, cost-efficient alternative that delivers near-frontier model performance with fewer compute resources.