Thinking Machines Unveils 276B Inkling-Small AI Model

Thinking Machines Lab has released Inkling-Small under an Apache 2.0 open-weight license. The multimodal reasoning model features a mixture-of-experts architecture with 276 billion total parameters and 12 billion active parameters per token. Despite reducing active parameters by roughly 71% compared to the original 975 billion parameter Inkling model—which used 41 billion active parameters—Inkling-Small scores 40 on the Artificial Analysis Intelligence Index, coming within one point of its predecessor.

The model matches or surpasses the full-sized Inkling on several benchmark tests, including achieving an 80.2% score on SWE-Bench Verified, alongside strong results on Terminal Bench and ARC-AGI-1. Inkling-Small natively processes text, image, and audio inputs across a context window extending up to one million tokens. Its decoder-only transformer uses 42 layers, alternating local and global attention. Inputs route through six of 256 selectable experts plus two shared experts, using a hierarchical patch encoder for visual data and a dedicated audio tokenizer.

Engineered via distillation from the main Inkling model followed by two weeks of reinforcement learning on NVIDIA GB300 NVL72 systems, the network can be fine-tuned using Tinker. When quantized to NVFP4, Inkling-Small fits onto a single NVIDIA B300 GPU with roughly 180 GB of VRAM. For practitioners, this dramatic reduction in compute requirements unlocks high-tier multimodal reasoning and agentic software engineering capabilities on standalone hardware without sacrificing enterprise-grade context lengths or open-source licensing flexibility.