Thinking Machines Lab Releases Inkling, Its First Open-Weight Multimodal Model

Thinking Machines Lab — the $12B AI startup founded by OpenAI alumni including Mira Murati and John Schulman — has released its first model, Inkling, a 975-billion-parameter open-weight system trained to handle audio, video, and text. The release marks a significant milestone for the company as it bids to become a serious player in an increasingly crowded AI race.

Thinking Machines Lab, the AI startup assembled largely from OpenAI defectors, has shipped its first model. Inkling is an open-weight, multimodal system trained from scratch to process audio, video, and text — and at 975 billion parameters, it sits firmly at the large end of the current model spectrum, requiring a cluster of specialized chips to run.

What Inkling Actually Is

The company is upfront that Inkling isn't setting benchmark records. But in a blog post accompanying the release, Thinking Machines describes the model as capable of advanced reasoning and coding, and says it delivers performance comparable to the best open-weight models currently available — which, notably, largely come from China.

One detail that stands out from the development process: during training, researchers noticed Inkling had quietly dropped natural language explanations for its chain-of-thought reasoning.

"It determined that the grammar was overhead, which is interesting," a company source told Wired.

The team reinstated natural language reasoning to keep the model's decisions interpretable — a meaningful choice for enterprise and safety-conscious use cases.

In another sign of AI's increasingly self-referential development loop, Inkling was used to fine-tune and improve itself during training.

Why Open-Weight Matters Here

The open-weight release is ideologically consistent with the company's stated mission. In a recent blog post, Thinking Machines argued that AI shouldn't be concentrated in the hands of a few gatekeepers — that more people should be able to build models on their own data, for their own use cases.

For startups and technical teams, the practical upside is real:

No per-token API fees — download and run it yourself
Modifiable weights — fine-tune for specific domains
Comparable performance to leading closed models at a fraction of the operational cost

Open-source and open-weight models have been gaining ground precisely because the economics are compelling, especially for companies building products that require heavy inference workloads.

The Team Behind It

Thinking Machines was founded in February 2025 by a roster of recognizable names from OpenAI:

Mira Murati — former CTO (and briefly interim CEO) of OpenAI
John Schulman — OpenAI cofounder, a key architect of ChatGPT
Lilian Weng — former VP at OpenAI, led safety and robotics research

The company raised what was described as the largest seed funding round in history, reaching a $12 billion valuation before shipping a single model. Prior to Inkling, the lab had released Tinker (a fine-tuning tool), a voice interaction prototype, and several research papers.

Competitive Context

Thinking Machines enters a market that is expensive, fast-moving, and increasingly defined by defector-founded companies. Anthropic, built by former OpenAI researchers including Dario and Daniela Amodei, recently filed for an IPO that could value it at over $1 trillion. Its model Claude has carved out a strong position in enterprise and coding use cases.

For founders and builders, the arrival of Inkling adds another capable open-weight option to the stack — one backed by a well-resourced team with deep model development experience. Whether Thinking Machines can sustain that trajectory without a massive commercial push remains the open question, but the model release at least confirms the lab is building, not just fundraising.