Thinking Machines Shrinks Inkling to 12B Active Parameters Without Giving Up Performance

Thinking Machines Lab has released Inkling-Small, an open-weight AI model that it says delivers performance comparable to its flagship Inkling model while requiring only a quarter of the active parameters. 

The model has 276 billion total parameters with 12 billion active parameters, compared with Inkling’s 41 billion active parameters, and supports a context window of up to one million tokens.

The company said Inkling-Small is available immediately through Tinker with a limited-time discount alongside the original Inkling model. “It’s a capable and efficient model, and an important stepping stone for us as we continue our work,” it said.

Inkling-Small was trained on NVIDIA GB300 NVL72 systems and supports native reasoning across text, images and audio. It is being released with full model weights and is also available for fine-tuning on the company’s Tinker platform and for multimodal chat through Tinker Playground.

“As we build our model family, we are always iterating on our approach,” the company said. It added that Inkling-Small benefited from changes to its pre-training data mix and training recipe, and that an earlier preview checkpoint was post-trained using on-policy distillation with Inkling as the teacher before further reinforcement learning for agentic coding.

This is the second model in the Inkling series. A few days ago, the company launched Inkling, a mixture-of-experts (MoE) transformer with 975 billion total parameters and 41 billion active parameters. The multimodal model supports a context window of up to one million tokens and was pre-trained on 45 trillion tokens spanning text, images, audio, and video.

According to Thinking Machines Lab, Inkling-Small matches or surpasses Inkling on several reasoning and coding benchmarks despite its smaller active parameter count. 

The company said the model scored 31.6% on Humanity’s Last Exam, ahead of Inkling’s 29.7%, and exceeded 80% on SWE-bench Verified. It also said Inkling-Small offers a better performance-to-compute trade-off across several reasoning and agentic tasks.

The model also retains multimodal capabilities. Thinking Machines Lab said Inkling-Small uses the same encoder-free architecture as Inkling, allowing it to process audio, images and text together. The company said it improved the model’s ability to use Python for visual tasks, including document and chart analysis, while maintaining performance across visual reasoning, speech understanding and longer-form audio reasoning.