Inkling downsizes without sacrificing its multimodal capabilities

Thinking Machines Lab's Inkling-Small model reduces its active parameters to 12 billion, an optimization that redefines the cost of reasoning.

Inkling-Small is a more compact open-weights version of Thinking Machines Lab's model. Its Mixture-of-Experts architecture combines 276 billion parameters in total, but only activates 12 billion for each query, compared to 41 billion for Inkling. The company is making the model's full weights available, without, however, presenting the project as open source.

The model natively accepts text, images, and audio, with a context window of up to one million tokens. Its reasoning level can be adjusted across five tiers, allowing the amount of compute used to be modulated based on cost, speed, or task difficulty.

Despite its reduced size, Inkling-Small outperforms its predecessor on several reasoning and tool-use benchmarks. It scores 31.6% on Humanity’s Last Exam without tools, compared to 29.7% for Inkling, as well as 80.2% on SWE-bench Verified and 64.7% on Terminal-Bench 2.1. The main model, however, retains an advantage in knowledge coverage and factuality.

This efficiency relies notably on a new training data composition, a revised learning recipe, and a distillation phase using Inkling as a teacher. Thinking Machines Lab also continued reinforcement learning dedicated to code and agentic tasks for two weeks.

Inkling-Small can be fine-tuned in Tinker or used in Tinker Playground for text, image, and audio interactions. Its training was conducted on NVIDIA GB300 NVL72 systems.