Thinking Machines Lab released Inkling-Small, a 12 billion active parameter model that delivers performance comparable to its larger Inkling model while using substantially less compute.

The mixture-of-experts transformer contains 276 billion total parameters with 12 billion active. Full weights are available on Hugging Face, with fine-tuning through Tinker and multimodal chat in Tinker Playground.

Inkling-Small features a one million token context window and variable thinking effort levels from minimal to xhigh. This allows developers to trade compute for performance across experimentation, coding, tool use, and production applications.

The model was trained on NVIDIA GB300 NVL72 systems using a revised pre-training data mix and machine learning recipe developed after Inkling's completion. An earlier checkpoint underwent post-training through on-policy distillation with Inkling as teacher, followed by two weeks of scaled agentic coding reinforcement learning.

Benchmark performance

Inkling-Small outperformed its larger sibling on reasoning and agentic coding tasks. The model achieved 31.6% on Humanity's Last Exam compared to Inkling's 29.7%, and reached 80.2% on SWEBench Verified.

Additional results include 55.9% on SWEBench Pro evaluation and 64.7% on Terminal-Bench 2.1. However, Inkling maintains advantages in knowledge coverage and factuality, according to the lab.

The model retains Inkling's natively multimodal, encoder-free design. Audio converts to dMel spectrograms while images split into 40-by-40-pixel patches processed through a four-layer hMLP before joining text tokens.

Inkling-Small can use Python for visual analysis, combining reasoning with cropping, zooming, and programmatic inspection of documents and charts. Safety post-training follows Inkling's recipe with 98.4% on StrongREJECT and 96.9% benign answer rate.

Both Inkling models are currently available on Tinker with limited-time pricing discounts.