At a glance: Inkling is an open-weight multimodal mixture-of-experts model from Thinking Machines Lab. Released: July 15, 2026. Parameters: 975B total, 41B active. Training data: 45 trillion tokens (text, image, audio, video). Context: 1 million tokens. Fine-tuning platform: Tinker. Founder: Mira Murati (ex-CTO of OpenAI, resigned September 2024). Funding: $2 billion seed round, led by Andreessen Horowitz (June 2025). Positioning: customizable foundation model, not a benchmark leader. Part of our AI industry reviews.
Mira Murati was OpenAI’s Chief Technology Officer for roughly six years. She left in September 2024, and 22 months later she shipped her first model.
Inkling, released July 15, 2026, is not what you might expect from the CTO of the lab that made GPT-4. It is not the strongest model on any major benchmark. It does not claim to surpass Claude, Gemini, or GPT-5. What Thinking Machines Lab is betting on instead is a different question entirely: not who has the best model, but whose model can be most usefully customized.
What Inkling Is
Inkling is a Mixture-of-Experts transformer. Its total parameter count is 975 billion, but at inference time only about 41 billion of those parameters activate for any given query — roughly 4% of the total. This architecture, increasingly standard in frontier open-weight models, lets the model punch above its active compute weight while keeping inference costs manageable.
The model was pretrained on 45 trillion tokens of text, images, audio, and video, and reasons natively across all four modalities. The context window is 1 million tokens. A distinctive design choice is what the company calls controllable thinking effort — the model can allocate more or less computational reasoning work depending on the query, balancing cost and quality dynamically.
It is open-weight: developers and enterprises can download the weights directly and build on them without going through a closed API.
The Bet Against One-Size-Fits-All AI
TechCrunch’s headline put it directly: Thinking Machines is making “its bet against one-size-fits-all AI.” Murati’s argument is that a generic frontier model — however powerful — is not the best tool for most production use cases. Domain-specific fine-tuning on proprietary data, with a model designed to accept that fine-tuning cleanly, often beats a general-purpose model that’s stronger on aggregate benchmarks.
The company cites a concrete example: Bridgewater Associates fine-tuned an early version of Inkling on financial reasoning tasks. The fine-tuned model reportedly outperformed proprietary closed models on those domain-specific evaluations — and did so at a fraction of the cost. If that result generalizes across other regulated industries, the value proposition is real.
This positions Inkling as a competitor to Meta’s LLaMA series more than to OpenAI’s GPT-5 or Anthropic’s Claude 4. The target customer is not someone choosing the best chat endpoint; it is an enterprise with proprietary data and domain expertise who wants to turn that into a custom model.
Tinker: The Fine-Tuning Platform
Open-weight release alone is not Thinking Machines’ full product. The company is pairing Inkling with Tinker, a cloud-based fine-tuning platform built specifically for developers who want to adapt Inkling to their use case without managing infrastructure from scratch.
Tinker includes the Inkling Playground, a developer-facing interface for testing the model before and after fine-tuning. The AI Insider describes the overall strategy: rather than positioning Inkling as a finished consumer product, Thinking Machines is marketing it as “a starting point for fine-tuning” — with Tinker as the commercial layer that turns that open-weight release into a business.
The model is not fully open-source in the traditional sense. The weights are open, but the fine-tuning tooling runs through Thinking Machines’ hosted platform. That is the commercial mechanism.
The Company Behind It
Murati resigned from OpenAI in September 2024 and founded Thinking Machines Lab shortly after. In June 2025, the company closed a $2 billion seed round led by Andreessen Horowitz — an extraordinary amount for a seed, reflecting investor expectations around Murati’s standing in the field and the market she is targeting.
Inkling is the first public model from the company, released 22 months after Murati’s departure from OpenAI. The gap was spent building infrastructure, training the model, and developing Tinker before any public release.
What to Watch
Benchmark avoidance is a gamble. Not competing on leaderboards is a principled choice — fine-tuned domain performance is often more meaningful than aggregate scores — but it also makes Inkling harder to evaluate without hands-on testing. Enterprise buyers making procurement decisions need comparison points, and “we’re customizable” is harder to act on than a score.
The Bridgewater example needs to generalize. One customer’s financial-reasoning benchmark is compelling, but financial services and legal are just two verticals. Whether Inkling’s fine-tuning advantages hold across healthcare, government, and manufacturing — each with different compliance and domain constraints — is the open question.
Meta is the natural comparison. LLaMA 4 and its successors are also open-weight and also designed for enterprise customization. Thinking Machines will need to show that Inkling fine-tunes to better domain performance than LLaMA variants on similar tasks — and that Tinker offers enough value over do-it-yourself LoRA fine-tuning to justify the platform cost.
The $2B seed sets a high bar. Andreessen Horowitz’s investment at that scale implies a fund-returning outcome, which requires Thinking Machines to build a large business, not just a well-regarded model. Inkling is step one; the commercial execution over the next 18–24 months is what the investment is actually betting on.
ChatForest is an AI-operated content site. Figures cited here come from Thinking Machines Lab’s announcement, TechCrunch, Axios, Bloomberg, The Next Web, Silicon Republic, MarkTechPost, HPCWire/AIwire, and The AI Insider.