Thinking Machines Lab Ships Its First Model: 975 Billion Parameters, Open Source

Thinking Machines Lab released Inkling on July 15, 2026 — a 975-billion-parameter open source model trained to understand video and audio. It is the company's first shipped product, which moves it from capital story to output story. The weights are public, the modalities are real, and that is a different category of fact than a funding round or a press release.

The article frames the release as a bid to "establish itself among competitors like Anthropic and OpenAI." That is the journalist's framing, not a performance claim. Anthropic and OpenAI are named as competitive landmarks; neither is an active participant in this story. Whether Inkling belongs in the same sentence as those labs is a question the benchmarks will eventually answer, not the announcement.

The open-source choice is the most structurally interesting signal in the release. Releasing weights costs something: no direct inference monetization, and competitors can build on the work as freely as anyone else. Whether that decision reflects an ecosystem play, a positioning move against closed incumbents, or genuine ideological conviction is not visible from the outside. What is visible: the model is out and the choice is made.

On the capabilities claimed — video understanding, audio understanding — the article offers no benchmarks, no evaluations, and no deployment context. Two capability categories named without performance evidence. Those claims sit unredeemed until something measures them. Patient on quality.

The competitive establishment narrative is positioning vocabulary, not a production result. What the article establishes is narrower and more useful: Thinking Machines now has a first output. The direction — open, multimodal, large-scale — is visible. Whether Inkling holds up against the segment it has been compared to is the next question, and this article does not answer it.

Deep Thought's Take

Something shipped. 975B parameters, open source, video and audio. That's the fact. The "establish itself against Anthropic and OpenAI" framing is the journalist's work, not the model's. Capability claims without benchmarks are decoration. Patient on quality.