Mira Murati Bets on Open Weights: Inkling vs. the Closed AI Stack

On July 15, 2026, Thinking Machines Lab published Inkling, a 975-billion-parameter open-weight model, releasing full weights under the Apache 2.0 licence and positioning it as a structural alternative to the proprietary APIs sold by OpenAI and Anthropic. The move crystallises a bet that former OpenAI CTO Mira Murati has been building toward since founding the company in early 2025: that enterprises will pay more for customisable infrastructure than for commodity intelligence delivered by the token.

The open-weight wager: 975 billion parameters, zero licensing fees

Inkling is a Mixture-of-Experts (MoE) model: 975 billion total parameters, of which 41 billion are active per token. The company trained it on 45 trillion tokens spanning text, images, audio, and video; the model processes all four modalities as input and returns text, with a 1-million-token context window. A smaller companion model, Inkling-Small, was previewed alongside the main release: 276 billion total parameters, 12 billion active per token.

Unlike recent flagship releases from OpenAI, Anthropic, and Google, Thinking Machines Lab made no claim to state-of-the-art performance on standard benchmarks. The company stated plainly at launch that Inkling "is not the strongest overall model available today," framing it instead as a purpose-built base for downstream fine-tuning, per TechCrunch. The full weights are on Hugging Face under Apache 2.0, with no licensing restrictions on commercial use or derivative works.

The company published a Bridgewater Associates case study alongside the release: a version of Inkling fine-tuned on Bridgewater's proprietary financial data scored 84.7 percent on internal financial-reasoning tests, outperforming the proprietary APIs it replaced, while costing roughly one-fourteenth as much to operate on an ongoing basis. That ratio is the core of Thinking Machines Lab's commercial argument.

Tinker as the real product: customisation over commodity AI

Thinking Machines Lab does not sell intelligence by the token. Revenue runs through Tinker: companies upload proprietary data, fine-tune Inkling or another open-weight base, and receive weights they can download or host on Tinker's infrastructure. The enterprise keeps the IP. Thinking Machines Lab charges a platform fee and takes a cut of hosted compute, not a per-token inference margin. That distinction matters structurally because inference margins have compressed sharply industry-wide since 2025.

StartupHub.ai data shows Thinking Machines Lab employs 212 people, compared to 10,452 at OpenAI - the smallest headcount-to-model ratio among the frontier AI labs we track. Among independent AI startups with released foundation models in our data, the nearest comparisons by size are Fei-Fei Li's World Labs (150 employees) and Sakana AI (199 employees), both similarly platform-oriented rather than operating large inference clusters. The contrast with OpenAI's headcount reflects a deliberate choice to build on open-weight infrastructure rather than train and serve proprietary models at hyperscale.

TechCrunch reported that the company has claimed a nine-month path from founding to first model, contrasting it with OpenAI's five-year and Anthropic's three-year trajectories from incorporation to a publicly released model. Murati outlined the company's human-AI collaboration thesis at Bloomberg Tech 2026 in San Francisco in early June 2026, her first major public appearance since founding Thinking Machines Lab, according to Bloomberg.

StartupHub.ai previously reported on how Tinker's architecture is designed to let enterprises replace general-purpose APIs with domain-specific models trained on their own data.

Nvidia's gigawatt vote of confidence

On March 10, 2026, Nvidia announced it would make a "significant investment" in Thinking Machines Lab and supply the startup with at least one gigawatt of next-generation Vera Rubin compute capacity under a multiyear agreement, according to Bloomberg. The financial terms were not disclosed. CNBC confirmed the announcement the same day.

One gigawatt of compute capacity is enough to power roughly 750,000 homes. Securing access to Vera Rubin systems at that scale substantially de-risks the compute bottleneck that constrains most frontier model training programmes, and it arrived before Thinking Machines Lab had publicly released a single model, underscoring how early the strategic commitment was made.

Bloomberg reported in November 2025 that Thinking Machines Lab was in funding discussions that had reached a $50 billion post-money valuation target. Those talks stalled by January 2026 without a disclosed close. The Nvidia partnership, announced two months later, provides an alternative form of capital: compute access tied to a strategic relationship with the dominant AI chip supplier, rather than a cash raise that would require a disclosed valuation at a difficult fundraising moment.

What it means

Murati is not attempting to out-GPT OpenAI. Inkling's explicit non-claim on benchmark leadership signals a structural choice: Thinking Machines Lab is competing on the fine-tuning and customisation layer, not the raw performance layer. If Tinker can convert a general-purpose open model into a domain specialist at one-fourteenth the ongoing cost of a proprietary API, the commercial case does not depend on topping any leaderboard. The risk is distribution - whether enterprise IT buyers will trust a 212-person lab's infrastructure over OpenAI's enterprise contracts or Anthropic's safety-forward positioning. The Bridgewater reference suggests the early answer is yes, at least in financial services, and the Nvidia deal gives the company enough runway to find out in others.

Sources

• Bloomberg: Murati's Thinking Machines Releases First AI Model for Broad Use
• TechCrunch: Thinking Machines amps up its bet against one-size-fits-all AI with Inkling
• Bloomberg: Nvidia to Invest in Mira Murati's Thinking Machines Lab and Supply Chips
• Thinking Machines Lab: Inkling: Our Open-Weights Model
• CNBC: Nvidia makes significant investment in Mira Murati's Thinking Machines Lab
• Bloomberg: Thinking Machines Murati on AI's Next Chapter (Bloomberg Tech 2026)

Editorial standards: every claim is sourced. Tips: [email protected]