Inkling Explained: Thinking Machines Lab's First Open-Weights Model

Inkling is the first AI model from Thinking Machines Lab. It launched on July 15, 2026, with open weights. That means anyone can download it and retrain it on their own data. Inkling uses a Mixture-of-Experts design with 975 billion total parameters but only 41 billion active per query.

The company was founded by Mira Murati, the former chief technology officer of OpenAI. Inkling is built to loosen the grip of closed frontier labs. It gives US teams an open-weights option to sit alongside Chinese open models like Qwen.

This guide explains what Inkling is in simple terms. You will learn how its architecture works, how it compares to models like GPT-5 and Claude, and how to customize it with Tinker. We also cover who should use it, its limits, and how to access it.

What Is Inkling?

Inkling is an open-weights foundation model released by Thinking Machines Lab on July 15, 2026. It is the company's first product-grade model. Open weights mean the model files are public, so anyone can run or modify them.

Thinking Machines was founded by Mira Murati, who was chief technology officer at OpenAI. The goal is to offer a strong, open alternative to closed models. Inkling is aimed at teams that want more control over their AI.

The company calls Inkling a broad, balanced foundation model. It says the model is strong across many domains and flexible enough to adapt. Thinking Machines was also clear that Inkling is not the strongest overall model available today, open or closed.

Inkling's Architecture and Open Weights

Inkling uses a Mixture-of-Experts (MoE) design with 975 billion total parameters and 41 billion active per query. In a MoE model, only a small part of the network wakes up for each request. Thinking Machines describes it as waking only a fraction of the brain to handle any query.

This design keeps Inkling cheaper and faster to run than its total size suggests. The model was pretrained from scratch on a multimodal mix of text, images, audio, and video. According to the company, it was trained entirely on state-of-the-art Nvidia hardware.

Because Inkling ships with open weights, teams can download it and fine-tune it on their own data. That is a key difference from closed models like GPT-5 or Claude, which you can only reach through an API.

• 975 billion total parameters, but only 41 billion active per query (MoE design).
• Pretrained from scratch on a multimodal mix of text, images, audio, and video.
• Trained on state-of-the-art Nvidia hardware, per the company.
• Open weights: anyone can download, run, and retrain the model.
• MoE routing keeps inference cheaper and faster than the full parameter count implies.

How Inkling Compares to GPT-5, Claude, and Qwen

Inkling trades raw power for a balance of cost and performance, unlike top closed models such as GPT-5 and Claude. Thinking Machines says Inkling is not the strongest overall model available today. Instead, it aims to be broad and well-rounded across many tasks.

Against closed models from OpenAI and Anthropic, Inkling's edge is openness. You can host it yourself, inspect it, and fine-tune it freely. Closed models offer strong performance but keep their weights private and run only on the vendor's servers.

Against Chinese open models like Alibaba's Qwen and startups such as Z.ai, Inkling offers a US-built open-weights choice. This matters for teams that prefer a domestic option for legal or supply-chain reasons. The trade-off is that Inkling may not top raw benchmark charts.

Customizing Inkling with Tinker

You customize Inkling through Tinker, Thinking Machines' cloud-based fine-tuning tool for AI developers. Tinker is the company's first product. It lets teams adapt Inkling to their own data and tasks.

Fine-tuning teaches the base model your domain, tone, and rules. A law firm could tune Inkling on its own briefs and policies. A clinic could tune it on approved medical language and workflows.

Because Inkling has open weights, you are not locked into one vendor. You can fine-tune with Tinker or run the weights on your own infrastructure. That flexibility is a core reason to choose an open-weights model.

Who Should Use Inkling?

Inkling fits teams that want control, customization, and a balanced cost-to-performance ratio. It is a strong pick for regulated businesses that need to host models in-house. It also suits developers who want to fine-tune a capable base model.

It is less ideal for teams that only need the single highest benchmark score. Those users may prefer a top closed model. Inkling shines when openness and adaptability matter more than raw peak power.

• Regulated SMBs that must keep data and models in their own environment.
• Developers who want to fine-tune a broad base model with Tinker.
• Teams seeking a US-built open-weights alternative to Chinese open models.
• Businesses balancing budget against performance, not chasing the top benchmark.

Limitations and Safety Considerations

Inkling is not the strongest overall model available today, by the company's own account. It is built for balance, so specialized closed models may beat it on some tasks. Buyers should test it against their real workloads before committing.

On safety, Thinking Machines said it tested Inkling for risks such as bio-weapons uplift and help with cyberattacks. The company reported that the model performed well in those tests. Open weights still raise a hard question: anyone can modify the weights, which can weaken safeguards.

Thinking Machines said it is still studying how open-weights safeguards can be tweaked over time. Teams that deploy Inkling should add their own guardrails and monitoring. Treat published safety results as a starting point, not a guarantee.

How to Access and Run Inkling

You access Inkling by downloading its open weights or by customizing it through the Tinker fine-tuning service. Because the weights are open, you can host the model on your own hardware. You can also run it through cloud providers that support open-weights models.

Running a 975-billion-parameter MoE model needs serious compute, so plan your infrastructure early. Many teams start with a fine-tuned version tuned for their exact use case. This keeps costs down and improves accuracy on your data.

If you are unsure where to begin, a short planning session can map the right setup. The goal is to match Inkling to your budget, data rules, and performance needs.