Thinking Machines Lab Releases Inkling-Small, a Highly Efficient Open-Weights MoE Model

Thinking Machines Lab has released Inkling-Small, a multimodal Mixture-of-Experts (MoE) open-weights AI model with 276 billion total (12 billion active) parameters. Despite being roughly one-quarter the size of the original Inkling model, it delivers comparable performance and even surpasses its predecessor in specific reasoning and agent tasks. This release demonstrates that smaller, optimized MoE models can achieve performance parity with much larger models, significantly lowering the computational barrier for advanced AI reasoning. By providing open weights and native multimodal capabilities, it empowers developers to run and fine-tune highly capable models more efficiently. Inkling-Small supports native audio and image reasoning, variable thinking intensity, and a 1-million-token context window, and was trained on NVIDIA GB300 NVL72 systems. It scored 31.6% on the "Humanity's Last Exam" benchmark, outperforming the original Inkling's 29.7%, and is available for fine-tuning via the Tinker API.