Mira Murati, the former chief technology officer of OpenAI, has delivered the first public product from her AI startup Thinking Machines Lab: a fully open-weight model named Inkling, designed to be customized by businesses and developers rather than used as a fixed, proprietary service.
Murati departed OpenAI in 2024, triggering widespread speculation about her next move. Thinking Machines subsequently raised capital and, according to earlier reporting, recruited more than 20 engineers from her former employer. Until now, the startup had disclosed little about its technical direction. The release of Inkling on July 16, 2026, is the first concrete signal of what that work produced.
What Inkling Is and How It Works
The model’s defining characteristic is the public availability of its weights — the numerical parameters that encode everything a large language model has learned during training. For most mainstream AI products, including OpenAI’s ChatGPT, those weights are proprietary and inaccessible. Users interact with the model through an API or interface and must, as the company itself frames it, “blindly trust” that the output aligns with their needs.
Thinking Machines is taking the opposite approach. According to the company, Inkling is “a model we trained from scratch with the full weights available, so that people can make it their own.” To make that customization practical, Thinking Machines has paired Inkling with a companion tool called Tinker, which allows users to adjust the model’s weights directly — effectively reshaping how the AI reasons and responds without requiring full retraining from scratch.
The company describes Inkling as “designed to be broad,” trained across what it lists as agentic tasks, reasoning, coding, instruction-following, factuality, vision, and audio — a deliberately wide scope. The rationale, according to Thinking Machines, is that real-world enterprise deployments require models that “can adapt to very different workflows,” rather than systems optimized narrowly to perform well on standardized benchmark tests.
That framing represents a subtle but pointed critique of how the frontier AI race has been fought. The leading labs — OpenAI, Anthropic, and Google DeepMind — have largely competed on headline benchmark scores, treating leaderboard performance as a proxy for real-world usefulness. Thinking Machines is arguing, implicitly, that benchmark optimization and practical adaptability are different things, and that enterprise buyers are starting to notice the gap. If correct, Inkling’s architecture could appeal to the segment of the market that has grown frustrated with paying for capable models that still require significant prompt engineering and external fine-tuning to fit specific workflows — a tension explored in broader industry commentary, including Satya Nadella’s warning that enterprises are paying for AI twice.
Thinking Machines’ stated mission is “to build AI that extends human will and judgment,” according to its website. The Tinker customization tool is positioned as the practical expression of that philosophy — the mechanism through which users exercise judgment over the model’s behavior rather than simply consuming outputs.
The open-weight approach also has broader policy resonance. A coalition that includes Nvidia, Microsoft, and roughly two dozen other technology companies has been actively lobbying U.S. lawmakers to support open-source AI development, arguing it accelerates innovation and reduces concentration risk. Inkling’s release adds a high-profile name to the open-weight camp at a moment when that debate is particularly active.
How Inkling Compares to Closed and Open Alternatives
| Model / Platform | Weights Access | Customization Depth | Primary Use Case | Benchmark Focus |
|---|---|---|---|---|
| Inkling (Thinking Machines) | Fully open (via Tinker tool) | Direct weight adjustment | Broad: agentic, reasoning, coding, vision, audio | Explicitly de-prioritized |
| ChatGPT / GPT-4o (OpenAI) | Closed, proprietary | Prompt engineering, fine-tuning API only | General-purpose; strong benchmark performance | High priority |
| Claude (Anthropic) | Closed, proprietary | System prompts, fine-tuning (enterprise tier) | Enterprise reasoning, safety-sensitive tasks | High priority |
| Llama (Meta) | Open weights (with use restrictions) | Full fine-tuning possible | Research, self-hosted enterprise deployment | Moderate priority |
The comparison above draws on publicly available information about each platform’s licensing and architecture policies. Inkling’s key differentiator within this landscape is the combination of open weights and a purpose-built consumer-facing customization tool, rather than requiring users to run their own fine-tuning pipelines. Meta’s Llama series offers open weights but targets developers comfortable managing infrastructure; Inkling is positioning Tinker as a more accessible on-ramp. Anthropic’s Claude Opus 5 remains closed-weight, despite its competitive pricing moves, meaning enterprise users still cannot inspect or modify the underlying model.
What This Means for the Industry
Inkling’s release forces a direct comparison with the two dominant paradigms in enterprise AI: the fully closed model (OpenAI, Anthropic, Google) and the open-weight model that requires significant developer overhead (Meta’s Llama). By pairing open weights with a guided customization interface, Thinking Machines is attempting to capture a middle market — businesses that want more control than a closed API allows but lack the infrastructure team to run self-hosted fine-tuning pipelines.
OpenAI has the most direct exposure here. Murati was one of the people who shaped GPT-4 and the products built around it; her public positioning of Inkling as a reaction against “blindly trusting” proprietary models is a credible critique from someone who operated inside that system. The company will need to demonstrate that its own customization offerings — prompt caching, fine-tuning APIs, enterprise system prompts — are sufficient for the segment Inkling is targeting.
Anthropic and Google DeepMind face a similar but less immediate pressure. Both have invested heavily in safety-aligned, closed architectures; open-weight models with deep customization introduce alignment risks those companies have historically cited as reasons to keep weights private. The debate over whether openness and safety are compatible will intensify as Inkling gains traction or fails to.
For the broader industry, the timing matters. The open-weight movement is gaining institutional credibility at exactly the moment when enterprise AI budgets are under scrutiny and buyers are asking harder questions about vendor lock-in. Thinking Machines has launched Inkling into that window deliberately. Whether the Tinker tool is robust enough to deliver on the promise of genuine, accessible customization — and whether Inkling’s performance across its stated task domains holds up against closed-model competitors in production settings — will determine how seriously the frontier labs are forced to respond.
