For eighteen months, Thinking Machines Lab has been one of AI's most-watched startups without a public product to show for it. Founded by Mira Murati, OpenAI's former chief technology officer, it raised a war chest, hired aggressively, and said almost nothing about what it was building.
That changed on Wednesday. And the way it changed says as much about the company's strategy as the model itself.
Thinking Machines didn't claim it had built the best AI model in the world. It said the opposite, in its own words, in its own launch post.
So What Actually Is Inkling?
Inkling is Thinking Machines' first in-house model, and it's open-weight, meaning anyone can download it, run it, and modify it directly, rather than only accessing it through a paid API like the flagship models from OpenAI, Anthropic, or Google.
A mixture-of-experts system with 975 billion total parameters, though only about 41 billion are active for any given task
Trained on 45 trillion tokens of text, images, audio, and video
A context window of up to 1 million tokens
Reasons across text, images, and audio natively, though outputs are currently limited to text, code, and structured data
Lets users dial "thinking effort" up or down to trade accuracy for speed
Ships alongside a preview of Inkling-Small, a lighter 12-billion-active-parameter version aimed at lower cost and latency
Full weights are available on Hugging Face, including a checkpoint tuned for NVIDIA's Blackwell chips, and the model is already accessible through hosting partners like Together, Fireworks, Modal, Databricks, and Baseten.
The Admission Everyone's Talking About
Here's the part that's turning heads. In its own release notes, Thinking Machines states plainly that Inkling is "not the strongest overall model available today, open or closed."
That's an unusual thing for any AI lab to say about its own flagship launch. Most releases lead with a benchmark chart and a claim about beating some rival. Thinking Machines led with a disclaimer.
The reason becomes clearer once you look at what the company is actually optimizing for. Inkling isn't pitched as a finished product you talk to. It's pitched as raw material, a broad, balanced base built specifically to be reshaped through Tinker, the company's fine-tuning platform released back in October 2025.
The argument, in plain terms: a model frozen at release loses out over time to one an organization can keep shaping around its own data and expertise. Inkling is designed to be good enough at everything to be a strong starting point, not to be the best at any one thing out of the box.
| Date | What Happened | What It Means |
|---|---|---|
| 2025 | Mira Murati founds Thinking Machines Lab after leaving OpenAI. | The company builds AI infrastructure largely out of public view. |
| October 2025 | Thinking Machines releases Tinker, its model-customization platform. | The company's first product ships before its first model does. |
| July 15, 2026 | Inkling and Inkling-Small are released, open-weight, on Hugging Face. | The company's first public model after 18 months of secrecy. |
| July 15, 2026 | Thinking Machines states Inkling isn't the strongest model available. | Signals the model is built for customization, not leaderboard wins. |
| Ongoing | Bridgewater Associates cites strong results fine-tuning an open model via Tinker. | An early proof point for the "customize, don't consume" pitch. |
Why This Matters Beyond One Company
Thinking Machines points to a project with hedge fund Bridgewater Associates as evidence the strategy works. Bridgewater used Tinker to fine-tune an open model, Alibaba's Qwen, on its own financial expertise, and reported a score of 84.7% on financial reasoning tests, beating leading proprietary models at a fraction of the cost. That figure comes from the two companies' own joint evaluation, not an independent benchmark, so it's worth treating as a case study rather than an industry-verified result.
The bigger backdrop here is a genuine gap in the market. Western open-weight models have fallen behind their Chinese counterparts in recent years, especially after Meta pulled back from its open strategy following Llama 4's rocky reception. That's left many businesses leaning on Chinese labs, including Alibaba's Qwen and DeepSeek, as their main affordable alternative to costly closed systems.
Microsoft CEO Satya Nadella has warned that companies relying on closed models effectively pay twice: once in usage fees, and again by handing proprietary knowledge from their prompts back to the model provider. Inkling is Thinking Machines' answer to that concern, a US-built, downloadable option that companies can adapt and keep in-house.
Inkling and Thinking Machines: FAQ
Inkling is Thinking Machines Lab's first in-house AI model, released July 15, 2026. It's an open-weight, mixture-of-experts model with 975 billion total parameters (about 41 billion active per task), trained on 45 trillion tokens of text, images, audio, and video.
No, and Thinking Machines says so directly: Inkling is "not the strongest overall model available today, open or closed." It's built as a broad, balanced base for customization rather than a leaderboard-topping model.
Developers and companies can download Inkling's trained parameters and run or modify it themselves, unlike closed models only accessible via API. Full weights are on Hugging Face, including a build for NVIDIA Blackwell hardware.
Tinker is Thinking Machines' fine-tuning platform, released in October 2025. Inkling is built to be customized through Tinker, letting companies fine-tune it on their own data and own the result, along with the safety responsibility that comes with it.
Thinking Machines points to Bridgewater Associates, which used Tinker to fine-tune an open model on its own financial expertise and reported an 84.7% score on financial reasoning tests. That figure is from the companies' own joint evaluation, not an independent benchmark.
Western open-weight models have lagged Chinese labs recently, pushing some enterprises toward models like Qwen or DeepSeek as their main affordable alternative to expensive closed systems. Inkling is positioned as a US-built open option in that gap.
Jans Bock-Schroeder
Publisher & Founder of AI Angst
Coming from the world of art, photography, and the luxury market, Jans launched AI Angst in 2025 to explore the cultural, ethical, and psychological impacts of artificial intelligence. His work bridges creative vision with critical technology analysis, offering clarity in an era of rapid technological change.
Sources and Citations
This article is based on the following primary sources, published July 15-16, 2026:
-
Thinking Machines Lab — "Inkling: Our open-weights model" (July 15, 2026)
Primary source for model specs, design philosophy, and the "not the strongest" self-assessment.
https://thinkingmachines.ai/news/introducing-inkling/ -
TechCrunch — "Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling" (July 15, 2026)
Source for benchmark context and the company's positioning strategy.
https://techcrunch.com/2026/07/15/thinking-machines-amps-up-its-bet-against-one-size-fits-all-ai-with-its-first-open-model-inkling/ -
Fortune — "Murati's Thinking Machines releases first AI model for broad use" (July 15, 2026)
Source for company valuation and competitive framing against OpenAI and Chinese labs.
https://fortune.com/2026/07/15/what-is-mira-murati-thinking-machines-first-ai-model-inkling/ -
Reuters (via Mezha) — "Thinking Machines unveils Inkling, a 975 billion parameter open weight model" (July 15, 2026)
Source for the Bridgewater/Tinker case study and open-source ecosystem context.
https://mezha.net/eng/bukvy/14a569a7_thinking_machines_unveils/ -
The Next Web — "Thinking Machines Inkling: Murati's first open-weight model" (July 15, 2026)
Source for the Bridgewater 84.7% benchmark figure and Satya Nadella quote context.
https://thenextweb.com/news/thinking-machines-inkling-open-weight-model
Published: July 16, 2026. Sources verified at time of publication. All external links open in a new tab.
