Inkling-Small matches its predecessor at roughly a quarter the size — and beats it on some benchmarks
Two weeks after releasing Inkling, Thinking Machines shipped a 276-billion-parameter multimodal reasoning model under Apache 2.0 that surpasses the 975B original on several benchmarks. Mira Murati's lab is arguing against one-size-fits-all frontier models by publishing the weights for both.
The size collapse is the story. Inkling is a natively multimodal mixture-of-experts system with 975 billion total and 41 billion active parameters, trained on 45 trillion tokens of text, images, audio and video, reaching 77.6 percent on SWE-bench Verified and 91.4 percent on VoiceBench. Inkling-Small reaches comparable ground at roughly a quarter of the parameters, exceeding the larger model on several measures.
Both ship under Apache 2.0 with a one-million-token context, which is a materially different proposition from the bespoke community licences attached to most large open-weight releases. Apache 2.0 at this capability tier removes the legal review that usually sits between an enterprise and a downloadable frontier model.
The lab also states an unusual design goal: Inkling was built to answer directly on topics subject to censorship, pitched at enterprises that want outputs determined by the facts rather than by sensitivity. That is a positioning claim rather than a measurable property, and it should be read as one.
Thinking Machines Lab — Introducing Inkling-Small → · Thinking Machines Lab — Inkling: our open-weights model → · VentureBeat — Thinking Machines debuts Inkling Small open source AI model nearing performance of predecessor at about 1/4 size → · TechCrunch — Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling →