Thinking Machines Lab says the weights for Inkling 2 Small are now available for download, a material change from its initial hosted release. The company added the November 12 update to its November 5 announcement, following the safety testing it had said would precede the weight release.
What changed
Our earlier report accurately described the initial release as hosted with weights still pending. The publisher now links the Hugging Face checkpoint, whose model card lists Apache 2.0. The company also identifies serving integrations including SGLang, vLLM, TokenSpeed and Unsloth, alongside hosted providers.
Inkling 2 Small is a mixture-of-experts model: the publisher describes 594 billion total parameters and 17 billion active per token, with image input. Active parameters describe computation per token; they do not mean the full model occupies only 17 billion parameters of storage. Review the model card and framework requirements before estimating hardware needs or costs.
Why the release matters
Available weights give developers more options to study, adapt and deploy a model than a hosted API alone. They do not guarantee that all training data, code or infrastructure are open, or that the model fits ordinary consumer hardware.
Performance figures in the publisher’s announcement are attributed claims. We have not downloaded the weights, run the model or independently reproduced its evaluations.
Original reporting from linked primary sources; no independent product or model test.
