Thinking Machines releases open-weight Inkling AI model

Thinking Machines ships first open model

Thinking Machines Lab released Inkling on Wednesday, marking the AI startup's first public model after 18 months of largely private development. Founded by former OpenAI CTO Mira Murati, the company is making a deliberate departure from the closed, general-purpose approach taken by OpenAI, Anthropic, and Google.

Inkling is an open-weight model, meaning organizations can download and modify it directly rather than access it only through API calls. The model uses a mixture-of-experts architecture with 975 billion total parameters, activating roughly 41 billion for any given task to balance performance with efficiency. It was trained on 45 trillion tokens spanning text, image, audio, and video, though current outputs are text-only, including code and structured data.

According to TechCrunch, which first reported the release, Thinking Machines explicitly states that Inkling is "not the strongest overall model available today, open or closed." Instead, the company is positioning it as a well-rounded foundation that enterprises can fine-tune through Tinker, its model-customization platform.

Why it matters

Thinking Machines is testing a fundamental thesis about enterprise AI: that models organizations adapt themselves will outperform standardized alternatives, even if those alternatives start with better benchmarks. This approach shifts safety and performance responsibility to customers — along with the requirement for serious machine-learning talent — but promises cost savings and domain-specific gains that closed models can't match. Microsoft CEO Satya Nadella recently warned that proprietary AI effectively charges enterprises twice: once in subscription fees and again by absorbing their business knowledge into future model versions.

Efficiency over dominance

The company claims Inkling uses one-third the tokens of Nvidia's Nemotron 3 Ultra to achieve equivalent coding performance. Users can adjust "thinking effort" to trade accuracy for speed, and the model is designed to flag uncertainty rather than guess.

A recent collaboration with Bridgewater Associates illustrates the potential upside. Researchers took an open-source model, trained it on Bridgewater's financial expertise, and reportedly achieved 84.7% accuracy on financial reasoning tests — outperforming top proprietary models while costing roughly one-fourteenth as much to run. Those results come from the companies' own evaluation, not independent testing.

Training and economics

Thinking Machines pre-trained Inkling from scratch but used outputs from other open-weight models, including Moonshot AI's Kimi K2.5, to generate early post-training data before reinforcement learning took over. The company says its next model will use fully self-contained post-training.

The startup trained Inkling entirely on Nvidia GB300 NVL72 systems through a partnership announced in March, but hasn't disclosed how it plans to cover those costs. A reported $50 billion fundraising round was said to be in progress last November but had stalled by January. Unlike OpenAI or Anthropic, Thinking Machines can't rely on metered API access for revenue — once weights are public, anyone can run them. The company's business model centers on Tinker, charging for training, fine-tuning, and a share of the hosting ecosystem.

Thinking Machines now employs roughly 200 people, up from levels reported after departures earlier this year, including two co-founders who left for OpenAI in January. The company says it reached market in about nine months, compared to roughly five years for OpenAI and three for Anthropic.

Details were first reported by TechCrunch.