Ex-OpenAI CTO Mira Murati’s startup, Thinking Machines Lab, has released Inkling, a 975-billion parameter open-weight AI model. Designed to challenge industry giants, it focuses on high customization and cost-to-performance over frontier benchmarks.
The new AI model fits into the enterprise landscape in several specific ways:
• Customization First: Inkling is an open-weight model, making it highly attractive to heavily regulated sectors like finance—such as early adopter Bridgewater Associates—that want to securely fine-tune models using localized or proprietary data without relying on generalized cloud APIs.
• Cost & Architecture: It utilizes a Mixture-of-Experts (MoE) setup where only about 41 billion parameters are active per query. This reduces computing overhead and latency, making it highly practical for complex, agent-based workloads.
• Multimodal Capability: Inkling natively handles text, images, and audio, and supports a context window of up to 1 million tokens.
Regarding concern about fully protecting trade secrets, this point is well taken. While open-weight models let enterprises fine-tune and run the AI on their own internal infrastructure (which significantly limits data leaks compared to public cloud APIs), the base models inherently learn patterns from their pre-training. Many high-security organizations—such as major defense or proprietary tech firms—still prefer completely isolated, local proprietary systems to guarantee 100% data privacy.
You can learn more about how the model is constructed and test its fine-tuning capabilities directly via the Thinking Machines Lab website.
“Former OpenAI technology chief Mira Murati is betting on more customizable artificial-intelligence models to chip away at the lead that frontier labs such as her former employer hold over the technology.
Thinking Machines Lab, the company led by Murati, released its first AI model on Wednesday -- and did it with "open weights," meaning others can modify it with their data. Called Inkling, the model has 975 billion total parameters, making it far smaller than estimates of the most advanced closed-source models from rivals such as OpenAI and Anthropic.
"We trained it for strong performance across the board rather than state-of-the-art in any single area, and it is not the strongest model available today, closed or open," the company said.
Thinking Machines's push into the decentralized ecosystem of open-weights AI models comes amid a broader industry backlash against the "walled garden" approach of frontier labs such as OpenAI and Anthropic.
Industry leaders such as Palantir Chief Executive Alex Karp and Microsoft's Satya Nadella have warned that companies, using OpenAI and Anthropic models, risk undermining their own business models by feeding their core institutional data into centralized, generalist models they don't control.
The release is also part of a push within Silicon Valley to build homegrown open-weights models as an alternative to those developed by China's Alibaba and a younger crop of startups such as Z.ai.
Many U.S. companies have been turning to Chinese open-weight models to help complete less-sophisticated AI tasks, in an effort to control costs and diversify their approach.
Rather than focus on raw power like the frontier labs, the Thinking Machines model was designed to balance "cost against performance," the company said. Of the nearly 1 trillion parameters that Inkling has, only 41 billion are "active," meaning that only a fraction of the AI's "brain" will be woken to deal with any query, making it cheaper and faster to use.
The model can be customized through Thinking Machines' first product, Tinker, a cloud-based [1] fine-tuning tool for AI developers and researchers released last year. The goal of Tinker is to allow a developer sitting at a laptop to customize and train large industrial AI models without having to worry about the supercomputing infrastructure underneath.
Last month, the hedge fund Bridgewater Associates and Thinking Machines released a report on Bridgewater's use of Tinker to fine-tune the Chinese open-weights model Qwen3-235B on its own data, leading to a model that Bridgewater said outperformed GPT-5 and Claude Opus on financial document triage, while cutting computing costs by over 13 times.
Thinking Machines pretrained it from scratch on 45 million tokens of text, images, audio and video. During post-training -- when the model is taught how to behave -- Thinking Machines used a combination of distillation, which relies on other AI models, and its own reinforcement learning process.
It was trained entirely on state-of-the-art Nvidia hardware. Thinking Machines and Nvidia announced a multiyear partnership in March in which Nvidia invested in the startup, which agreed to deploy at least one gigawatt of cutting-edge chips to train and serve its frontier AI models.
Thinking Machines said it tested the model for safety, including for risks such as the potential for it to help build biological weapons or aid hackers with cyberattacks, and it performed well. The company said it is still studying how safeguards it built into the model can be tweaked because it is open weights -- a safety concern raised in part by some proprietary model developers.
On Friday, Thinking Machines released its first manifesto, outlining its vision for a future in which AI was decentralized and built on local knowledge.
The company, whose CEO, Murati, witnessed the collapse of communism in her native Albania as a child, compared the current dominant AI paradigm of close-source frontier labs to "central planning" -- great for bounded tasks like chess and math, but not for the real work humans do every day.
"Central planning fails not because of insufficient intelligence, but because of the nature of productive knowledge: tacit, local, fleeting, and held privately by those who acquired it through their work," the company wrote, citing Friedrich Hayek.
"Attempting to aggregate knowledge for the use of a centralized intelligence faces the same challenge."” [2]
1. “The model can be customized through Thinking Machines' first product, Tinker, a cloud-based fine-tuning tool for AI developers and researchers released last year.” Does this cloud-based use increase risk for user's trade secrets to be stolen?
Whether a cloud-based tool like Thinking Machines' Tinker increases the risk of your trade secrets being stolen depends entirely on your risk tolerance and the specific cloud controls you implement.
Using managed cloud services introduces certain inherent risks compared to on-premise solutions:
• Data Transfer: Your proprietary fine-tuning data and algorithms must be sent to and processed on Thinking Machines' servers.
• Third-Party Exposure: You rely on the platform’s underlying security infrastructure (like Google Cloud) and the provider's data-handling policies to protect your information.
However, the Tinker API is designed specifically to mitigate these risks and cater to researchers and enterprise developers who require strict data governance:
• Algorithmic Control: Instead of uploading a proprietary dataset to a "black-box" service, you define the training recipes and algorithms yourself in standard Python code.
• Selective Weight Updates: Tinker employs Low-Rank Adaptation (LoRA), which means you are generally only training and sending streamlined adapter weights (a fraction of the model) rather than retraining the entire base model with your raw data.
• Ownership: You maintain control over your fine-tuning data and your custom models, rather than giving the platform rights to absorb your business logic into their general model.
To evaluate whether this tool is viable for your organization, you have to weigh these built-in controls against your internal data security policies.
1. Thinking Machines Releases AI Model --- First model from ex-OpenAI executive's startup aims to cut the giants' lead. Keach Hagey; Berber, Jin. Wall Street Journal, Eastern edition; New York, N.Y.. 16 July 2026: B5.
Komentarų nėra:
Rašyti komentarą