Thinking Machines releases Inkling Small, matching its flagship at a quarter of the compute

Thinking Machines Lab, the AI startup founded by former OpenAI chief technology officer Mira Murati, released Inkling-Small on Wednesday — a 276-billion-parameter open-source model that comes within a single point of its 975-billion-parameter predecessor on the Artificial Analysis Intelligence Index, while running on just 12 billion active parameters per token.

The release follows Inkling's debut just two weeks earlier, making Thinking Machines one of the fastest labs to ship an efficient variant of a flagship model at near-parity performance. Inkling-Small uses a Mixture-of-Experts (MoE) architecture — only a subset of its 276 billion total parameters activates for each forward pass, reducing compute cost and inference latency without the full performance penalty of a genuinely smaller dense model.

What Inkling Small can do

The model accepts text, image, and audio inputs and generates text output. It supports a context window of up to one million tokens — matching Inkling — and features the same variable thinking effort mechanism that lets developers trade response speed against reasoning depth. Inkling-Small was trained on NVIDIA GB300 NVL72 systems and scores competitively against other open-weight models at the 12-billion active parameter level on Terminal-Bench 2.1 (agentic tool use), HLE (reasoning), and IFBench (instruction following).

The full weights are available on Hugging Face under an Apache 2.0 license, which permits commercial use, modification, and redistribution without royalties. Thinking Machines has also added fine-tuning support through its Tinker API and is running a 50% API discount at launch through the Tinker Playground interface.

Why enterprises should pay attention

The original Inkling at 975 billion total parameters (41 billion active) requires serious GPU infrastructure. Inkling-Small brings the active parameter count down to 12 billion per token — a 70% reduction in per-inference compute — while preserving much of the flagship's coding, reasoning, and multimodal performance. The model is still too large for a consumer workstation, but it fits meaningfully within reach of enterprises that have some GPU capacity without running hyperscale infrastructure.

The compressed gap between 975 billion and 276 billion parameters, with benchmark scores that barely move, also signals a broader efficiency trend in AI development: labs are increasingly proving that MoE architectures can deliver flagship-class capability at significantly lower serving cost. That dynamic puts pressure on both closed API providers and other open-weight labs. Thinking Machines was founded in early 2025 by Murati after her departure from OpenAI and has attracted significant venture backing.