Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has released Inkling, an open-weights AI model designed for customizable, domain-specific workflows, signaling a strategic pivot away from the proprietary model approach that has dominated Silicon Valley's AI investment boom. The move comes as the industry grapples with a mounting profitability crisis: companies have invested $3 trillion since ChatGPT's 2022 debut, with another $7.5 trillion planned over the next five years, yet the revenue needed to justify these costs keeps climbing.

Why Is Open-Source AI Suddenly Threatening Big Tech's Dominance?

For years, closed proprietary models from companies like OpenAI and Anthropic held a decisive performance advantage. But that gap is narrowing rapidly. In July 2026, Kimi K3, an open model from Chinese lab Moonshot AI, nearly matched the performance of Anthropic and OpenAI's flagship models, Fable 5 and GPT-5.6 Sol, which had been released just weeks earlier. Moonshot then released K3's weights publicly, making it the largest and most powerful open model available at 2.8 trillion parameters.

This shift matters because it undermines the business model that justified the infrastructure boom. The industry's bet has always been straightforward: people will pay subscriptions or see ads for AI services powered by proprietary models running in massive data centers owned by a handful of firms. But if open models can match closed ones, that calculus breaks down. Companies can download an open model, customize it for their own needs, and run it on their own servers, cutting out the middleman entirely.

What Makes Thinking Machines Lab's Approach Different?

Thinking Machines Lab, founded last year by Mira Murati after she left OpenAI, has embraced this open-model strategy from the start. The lab released Inkling as its first major product, followed just two weeks later by Inkling-Small, a leaner variant that achieves near-identical performance at a fraction of the size.

Inkling-Small demonstrates the efficiency gains possible with modern AI design. The model contains 276 billion total parameters but only 12 billion active parameters at any given time, thanks to a mixture-of-experts architecture that activates only the relevant parts of the network for each task. Despite its smaller footprint, Inkling-Small actually surpassed the full Inkling model on some reasoning and coding benchmarks, according to the company. Both models handle text, images, and audio natively, and both ship through Tinker, Thinking Machines Lab's training and inference platform designed to simplify fine-tuning for developers.

The company is offering both models at a 50 percent discount for a limited time, making them accessible to teams exploring open-weight alternatives.

How to Evaluate Open-Weights Models for Your Organization

• Performance Benchmarks: Compare models on reasoning tasks, coding challenges, and domain-specific benchmarks relevant to your use case, not just general knowledge tests. Inkling-Small's performance on agentic-coding tasks demonstrates that smaller models can excel in specialized areas.
• Customization Capability: Assess whether the model supports efficient fine-tuning through techniques like LoRA adapters, which allow you to adapt the model to your specific data without retraining from scratch. Tinker's hosted API for distributed fine-tuning exemplifies this approach.
• Infrastructure Requirements: Evaluate whether you can run the model on your own servers or in your preferred cloud environment, versus being locked into a vendor's proprietary platform. Open-weights models offer flexibility that closed models cannot match.
• Total Cost of Ownership: Factor in not just subscription fees but also the cost of infrastructure, fine-tuning, and maintenance. Smaller models like Inkling-Small may offer better economics for many tasks.

Is This the Beginning of the End for Big AI's Data Center Bet?

The broader context makes Thinking Machines Lab's timing significant. State and local governments have passed approximately 275 temporary or permanent data-center bans just in 2026, with another 75 in the works, according to reporting cited in industry analysis. Public sentiment is also shifting; recent polls show Americans have become significantly less trusting of AI and more worried about its societal impact, particularly among young adults.

Meanwhile, companies that have already invested heavily in AI are pumping the brakes. Some firms blew through their annual AI budgets in just over a quarter and are now demanding better metrics on return on investment. Others worry that AI companies will use their internal data to build competing products.

Small models represent another threat to the centralized data-center model. A startup called PrismML has demonstrated that it can compress larger models to fit on iPhones without catastrophic performance loss. Apple, which has deliberately stayed out of the data-center arms race and aims to keep AI processing on-device whenever possible, is reportedly in talks with PrismML.

"Imagine a world, maybe three years from now, where 95 percent of the intelligence that you need is available to you locally, on your phone, on your laptop, on your appliances, and it's really on the last maybe 5 percent of high-end stuff that you'll need to go to the cloud," said Babak Hassibi, CEO of PrismML.

Babak Hassibi, CEO at PrismML

This vision would fundamentally reshape AI economics. Most people use AI for straightforward tasks like searching the internet, summarizing documents, or creating presentations. Open and small models are already nearly as capable as closed models for these common use cases. The frontier improvements that justify massive infrastructure investments are concentrated on complex tasks like advanced coding and mathematical reasoning, which represent a minority of actual user demand.

Thinking Machines Lab's rapid iteration cycle underscores how quickly the open-model landscape is evolving. Inkling-Small benefited from improved pre-training data, on-policy distillation using Inkling as a teacher, and reinforcement learning from agentic coding, all accomplished in just two weeks. This pace suggests that the performance gap between open and closed models may continue to narrow faster than many in Silicon Valley anticipated.

The industry's historic $1.5 trillion investment bet remains predicated on a specific outcome: that proprietary models will remain so superior that users will pay premium prices for them. Mira Murati's pivot to open models, combined with the emergence of competitive open alternatives from Moonshot, Meta, and others, suggests that outcome is no longer guaranteed. Whether open and small models can truly capture the majority of AI demand remains to be seen, but the trajectory is unmistakable.