July 16, 2026, (Inside AI) — Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, launched its first in-house AI model on Wednesday. Called Inkling, it is an open-weight, 975-billion-parameter large language model that lets users dial thinking effort up or down.
Inkling is not the strongest model on the market, the company admits. But its mix of controllability, multimodal training, and open licensing marks a deliberate break from the one-size-fits-all approach of frontier labs. The model is available for download on HuggingFace under an open-source license.
Thinking Machines positions Inkling as a base for customization. Enterprises can fine-tune it through Tinker, the startup’s model customization platform. Revenue will come from that hosting ecosystem, not metered API access—a direct contrast to rivals like OpenAI and Anthropic.
The Architecture of Choice #
Inkling uses a mixture-of-experts design. It draws on only 41 billion of its 975 billion parameters per task, making it faster and cheaper to run. The model supports a context window of one million tokens.
Researchers trained it from scratch on 45 trillion tokens spanning text, image, audio, and video. It reasons natively across all four modalities, though outputs are currently limited to text—code, styled artifacts, or structured data.
During pre-training, the team used other open-weight models like Moonshot AI’s Kimi 2.5 to generate early post-training datasets. The next model will use fully self-contained post-training data, the company said.
Training ran entirely on Nvidia GB300 NVL72 systems, thanks to a March deal for a gigawatt of Vera Rubin computing capacity.
Performance and Positioning #
On coding benchmarks, Inkling matched Nvidia’s Nemotron 3 Ultra while using a third of the tokens. It also works inside various coding and agent harnesses, with a controllable thinking feature that trades speed for depth.
“Inkling is not the strongest overall model available today, open or closed,” the company stated. “Instead, a combination of qualities makes it a good open-weights base for customization: multimodal capabilities, efficient thinking, and availability on Tinker for fine-tuning.”
Alongside the main model, Thinking Machines previewed Inkling-Small, a lighter version with 12 billion active parameters. It promises strong performance at even lower cost and latency.
The launch comes as open-source LLM development gains traction in the US, partly driven by White House restrictions on cutting-edge closed models. Thinking Machines now employs roughly 200 people, up from earlier levels after a wave of departures, including two co-founders who left for OpenAI in January.
By betting on adaptable, open-weight models, Murati’s venture challenges the prevailing business models of major AI labs. But it also shifts safety responsibilities to the organizations that fine-tune Inkling—a trade-off that will test enterprise appetite for control over convenience.