Mira Murati's Thinking Machines ships its first model — open-weight, and deliberately not the strongest

More than a year after former OpenAI CTO Mira Murati founded Thinking Machines Lab, the startup released its first model. Inkling is an open-weight mixture-of-experts system — 975B total parameters with about 41B active per task — trained on 45 trillion tokens spanning text, image, audio and video, and reasoning natively across all four. Developers can download and customize the weights (training data and source code stay closed). The company's own framing is unusual: Inkling benchmarks well against comparable open models, but "it is not the strongest model available today, closed or open" — the pitch is adaptability against one-size-fits-all frontier labs.

Why it matters

The most-funded startup in AI history just revealed its actual strategy, and it isn't a frontier-lab clone: it's open weights plus per-customer tuning, aimed at the gap between closed APIs and raw open models. That lands in a suddenly crowded week — Moonshot's Kimi K3 is pushing open weights from the Chinese side — and it puts pricing and customization pressure on the closed incumbents already fighting a price war. US open-weight releases at this scale have been rare since Meta's retreat; Inkling is the first serious American entry in a lane China has been winning by default.

What to watch

Real-world adoption (fine-tune volume, enterprise deals) rather than benchmark placement, and whether Thinking Machines' revenue model — customization services on open weights — converts its ~$2B of raised capital into a business.