Thinking Machine Labs, founded by former OpenAI CTO Mira Murati, has released its first in-house AI model called Inkling, marking a significant bet on open-weight artificial intelligence as an alternative to proprietary systems from OpenAI and Anthropic. The model contains 975 billion total parameters, with 41 billion active parameters for any given task, and was trained from scratch on 45 trillion tokens spanning text, images, audio, and video.

What Makes Inkling Different From Closed AI Models?

Inkling's defining characteristic is that it is open-weight, meaning developers and companies can download the model and run it locally on their own hardware rather than relying on cloud-based APIs. This contrasts sharply with flagship models from OpenAI (like GPT 5.6) and Anthropic (Claude Fable), which remain fully proprietary and accessible only through paid cloud services. The open-weight approach makes Inkling significantly cheaper to operate, a critical advantage as organizations worldwide reassess AI spending due to rising costs.

Mira Murati announced the release on X, stating: "Our first model, Inkling. Trained from scratch, weights are open, fine-tunable on Tinker today." Tinker is a proprietary tool developed by Thinking Machine Labs that allows developers to customize Inkling for specific tasks and workflows.

Mira Murati

The company explicitly positions Inkling not as the strongest overall model available, but rather as the best open-weight foundation for customization. Thinking Machine Labs argues that AI systems organizations can shape for themselves will perform better than centrally trained, one-size-fits-all models. The model is designed to provide calibrated answers, signal uncertainty rather than guess, and allow users to adjust "thinking effort" to balance speed and performance.

How Does Inkling Perform Against Competitors?

On Terminal Bench 2.1, a widely used AI performance benchmark, Inkling matched Nvidia's Nemotron 3 Ultra model while using roughly one-third as many computational tokens, demonstrating efficiency gains. However, Inkling trails more advanced closed models including GPT 5.6, Claude Fable 5, and Moonshot AI's Kimi K2.6 on peak performance metrics.

Thinking Machine Labs confirmed that Inkling's post-training phase partially used outputs from other open-weight models, including Moonshot AI's Kimi K2.5, before large-scale reinforcement learning took over. The entire model was trained on Nvidia GB300 NVL72 systems under a partnership announced in March.

How to Deploy and Customize Inkling for Your Organization

• Download and Run Locally: Unlike proprietary models, you can download Inkling's full weights and run the model on your own servers or hardware, eliminating dependency on third-party cloud APIs and reducing operational costs.
• Fine-Tune Using Tinker: Thinking Machine Labs provides Tinker, a dedicated tool for fine-tuning Inkling to your specific use cases, workflows, and domain-specific tasks without requiring extensive machine learning expertise.
• Adjust Reasoning Effort: Users can calibrate the model's "thinking effort" parameter to balance speed and accuracy, allowing organizations to optimize for either faster responses or higher-quality outputs depending on their needs.

The company also released a preview of Inkling-Small, a lighter-weight variant with 12 billion active parameters designed to deliver strong performance with even lower computational costs and latency. Thinking Machine Labs plans to release additional models in the Inkling family in the future.

Why Does Open-Weight AI Matter in the Broader AI Landscape?

Inkling's release aligns with a growing philosophical debate about AI's future. David Siegel, co-founder of Two Sigma and chairman of the Siegel Family Endowment, recently argued in Fortune that the 1980s debate over whether software should be proprietary or shared is replaying at frontier-model scale, with higher stakes for science, medicine, and public infrastructure. Siegel noted that open-source software like GCC and GNU/Linux became load-bearing infrastructure precisely because transparency enabled worldwide developer communities to find and fix flaws.

Thinking Machine Labs' philosophy echoes this argument. The company believes AI should be decentralized rather than controlled by a few firms. Microsoft CEO Satya Nadella has similarly argued that enterprises using proprietary models face a double cost: subscription fees plus the surrender of business knowledge through the prompts and corrections they input.

However, Siegel also emphasizes a critical distinction: open-weight releases do not automatically mean full transparency. While Inkling's weights are available for download, the full training pipeline, data curation process, and reinforcement learning methodology remain proprietary. An explanation of how a model works is not the same as an audit of how it was trained, Siegel argued, noting that "a model's stated reasons are a plausible story assembled after the fact, not a faithful record of the computation that produced the answer".

Siegel

Inkling's launch represents Thinking Machine Labs' first major public product after roughly 18 months of building AI infrastructure largely out of public view. The startup, founded in February 2025, received approximately 12 billion dollars in seed funding at the time, the largest seed round in history. The company currently employs around 200 people.

The release comes as the broader AI industry shows a split screen: open-weight models like Inkling, Grok Build (Apache 2.0), and Gemma 4 are shipping with community updates, while the hardest agentic benchmarks and most advanced reasoning capabilities remain behind closed APIs controlled by a handful of companies. Siegel's core claim is that the trend toward closure at the frontier is accelerating faster than the open-weight counterweight is maturing, particularly as science and public infrastructure increasingly depend on AI systems.