Is Inkling-Small free to use?

Yes — it is released under an Apache 2.0 license with weights on Hugging Face, so it can be downloaded, modified and used commercially at no cost, though running it still requires substantial GPU hardware.

Can Inkling-Small run on a regular computer?

No. Even the compressed NVFP4 version needs at least 180GB of GPU memory — about two Nvidia H200 cards — far beyond consumer hardware.

How is Inkling-Small different from the original Inkling model?

Inkling-Small uses about a quarter of the active parameters (12 billion vs. 41 billion) but scores higher on reasoning (Humanity’s Last Exam) and coding (SWE-Bench Verified) benchmarks, though the larger Inkling still leads on general knowledge.

Thinking Machines Releases Inkling-Small, Outperforms Larger Model

Thinking Machines Lab, the AI startup founded by former OpenAI chief technology officer Mira Murati, released Inkling-Small on July 30 — two weeks after shipping its first open-weight model, Inkling. The new model needs a fraction of the compute per response yet matches or beats its bigger sibling on several benchmarks, according to the company’s published model card.

Smaller model, comparable results

Inkling-Small is a sparse Mixture-of-Experts transformer with 276 billion total parameters, but only 12 billion are active for any given token — about a quarter of the 41 billion active parameters in the original Inkling. Each token routes through 6 of 256 experts plus two shared experts across 42 decoder layers, and the model accepts text, images and audio as input while producing text output with visible reasoning tokens.

Despite the smaller active footprint, Inkling-Small scored 31.6% on the Humanity’s Last Exam benchmark versus Inkling’s 29.7%, and 80.2% on SWE-Bench Verified versus 77.6% for the larger model, Thinking Machines Lab reported. The company said Inkling still leads on raw knowledge coverage and factual recall, where its larger parameter count helps. On the StrongREJECT safety evaluation, which checks whether a model refuses clearly harmful requests, the two models scored within half a percentage point of each other.

Still not a laptop model

The model carries a permissive Apache 2.0 license, with weights published on Hugging Face and support for fine-tuning through the company’s Tinker platform. It supports context windows of up to 1 million tokens. But “small” is relative: the full-precision checkpoint still needs more than 600GB of combined GPU memory, and even the quantized NVFP4 build requires at least 180GB — roughly two Nvidia H200 cards. Thinking Machines Lab trained the model on Nvidia’s GB300 NVL72 systems.

The release continues a pattern for the startup Murati launched in 2025 with several other former OpenAI researchers: pairing frontier-level benchmark results with permissively licensed weights instead of closed APIs, a strategy the company has said is meant to “extend human will and judgment.”

Smaller model, comparable results

Still not a laptop model

Frequently asked questions