Artificial Intelligence

A Frontier Lab Admits Its Model Is Not the Best

Mira Murati's Thinking Machines released Inkling, a 975B-parameter open-weight model, and stated plainly that it is not the strongest model available. That sentence is the product decision worth reading.

Artificial IntelligenceBy Servola Tech Desk2026-07-166 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

A Frontier Lab Admits Its Model Is Not the Best

Key takeaways

Thinking Machines Lab released Inkling on 15 July 2026, a 975-billion-parameter open-weight model with 41 billion active parameters, and stated in its own announcement that it is not the strongest overall model available today. No frontier lab writes that sentence by accident.
The claim it makes instead is adaptability: Inkling is offered as a base to customise rather than a benchmark to admire, with weights published on Hugging Face and fine-tuning through the company's Tinker platform.
Its architecture largely follows DeepSeek-V3, and early post-training data was generated using open-weight models including Moonshot AI's Kimi K2.5. A lab founded by OpenAI's former chief technology officer built on Chinese open-weight foundations.
This changes the procurement question. 'Which model scores highest' is answered by a leaderboard that moves every few weeks. 'Which model can we make ours, and keep' is answered by a licence and a copy of the weights, and the second answer lasts longer.
Be honest about the cost before you commit. Customisation needs training data you own, an evaluation set that reflects your actual work, and a named person responsible for it. Without those three, an open-weight base is a download, not a capability.

The sentence a frontier lab does not usually write

Buried in Thinking Machines Lab's announcement of Inkling on 15 July is a line no other frontier lab has been willing to publish. The model, the company wrote, is "not the strongest overall model available today." It went on to describe what it thinks Inkling is instead: "a good open-weights base for customization." Every model launch of the past three years has led with a benchmark chart. This one led with a limitation.

Read that as strategy, not modesty. Mira Murati, formerly chief technology officer at OpenAI, founded Thinking Machines and took roughly nine months to ship a frontier-scale model, against something closer to five years for OpenAI and three for Anthropic. The lab employs around 200 people and lost two co-founders to OpenAI in January. A company in that position does not win a benchmark race against labs with a hundred times the compute budget. So it declined to enter one.

The useful part for anyone buying AI is that an honest limitation is information you can plan with. A vendor claiming the highest score on a public leaderboard is telling you about a number that will be beaten within weeks, quite possibly before your procurement process finishes. A vendor telling you what its product is not has given you something durable. That is rarer than it should be, and it is worth more than it looks.

What is actually in the box

Inkling is a mixture-of-experts transformer with 975 billion total parameters and 41 billion active at any one time. It was pretrained on 45 trillion tokens spanning text, images, audio and video, and supports a context window of up to 1 million tokens. It reasons across text, images and audio, though it currently writes back only text, including code and structured data. A lighter variant, Inkling-Small, carries 276 billion total parameters with 12 billion active.

Two design choices tell you what it is optimised for. The model exposes controllable thinking effort, letting a caller trade reasoning depth against token spend on a per-request basis, which is a cost lever most APIs do not hand you. Thinking Machines also chose relative positional embeddings over RoPE, reporting that they extrapolate better to longer sequences. The company says Inkling uses roughly one third as many tokens as Nvidia's Nemotron 3 Ultra for equivalent coding performance. On a metered bill, token efficiency is not a technical detail.

Check the practical numbers before the headline ones. The weights are on Hugging Face, and Inkling is available through Tinker, the company's fine-tuning platform, at context lengths of 64,000 and 256,000 tokens, currently with a 50% discount for a limited period. Note the gap: the model supports up to a million tokens of context, while the managed platform currently offers a quarter of that. If your use case depends on the full window, you are self-hosting it, and self-hosting a 975-billion-parameter model is a serious infrastructure commitment rather than an afternoon.

The open frontier is a shared inheritance now

The provenance detail in this release deserves more attention than the parameter count. Thinking Machines states that Inkling's mixture-of-experts design largely follows DeepSeek-V3, using 256 routed experts and two shared experts with six routed experts active per token. It also used open-weight models, among them Moonshot AI's Kimi K2.5, to generate data for early post-training before large-scale reinforcement learning. A San Francisco lab founded by OpenAI's former chief technology officer built its architecture on a Chinese lab's published design and bootstrapped it with another Chinese lab's model.

This is what an open frontier actually looks like in practice. Not a flag-planting race between two blocs, but a shared technical inheritance that flows in every direction and does not ask permission at a border. The engineering moved from Hangzhou and Beijing to San Francisco because it was published, and the result was published in turn.

For a European owner that has a specific and unsentimental value. Access to a hosted frontier model is a commercial relationship subject to policy, pricing and availability decisions taken in another jurisdiction, as anyone whose model launched in 47 countries but not the European Union has already learned. A downloaded set of open weights is a different kind of asset. It does not get restricted, repriced or deprecated out from under you, because you already have it. That is not an argument that open weights are better. It is an argument that they fail differently, and knowing how a dependency fails is most of what procurement is for.

Whether this is a purchase you should be making

The line to draw is between renting intelligence and owning a capability. If your AI work is broad, varied and changes weekly, rent it: pay per token for the best hosted model and switch when a better one appears, because you are buying flexibility. If your work is narrow, repetitive, high volume and specific to your business, a customisable base you fine-tune and keep can beat a general model that has never seen your domain, and its cost stops scaling with your usage.

Be honest about what customisation requires, because this is where the plan usually breaks. Fine-tuning needs training data you actually own and are permitted to use. It needs an evaluation set built from your real work, not a public benchmark, or you will have no way to know whether your version is better than the thing you started with. And it needs a named person who owns the model in production, because a fine-tuned model that nobody maintains degrades quietly as your business changes around it. Missing any one of those three, an open-weight base is a download, not a capability.

The honest read on Inkling is that the admission is the feature. Thinking Machines has told you what it is not, which means the only question left is whether adaptability is what you need. For most companies running a handful of general AI tasks, it is not, and renting remains correct. For the smaller number with a narrow high-volume problem and real proprietary data, a base you can own and shape is a genuinely different economic proposition than metered access to someone else's model. Both are defensible. Buying the second while budgeting for the first is not.

Frequently asked questions

What is Inkling and who made it?

Inkling is the first open-weight model from Thinking Machines Lab, the company founded by Mira Murati, formerly chief technology officer at OpenAI. Released on 15 July 2026, it is a mixture-of-experts transformer with 975 billion total parameters and 41 billion active, pretrained on 45 trillion tokens of text, images, audio and video, supporting up to 1 million tokens of context.

Why would a lab say its own model is not the best?

Because it is competing on a different axis. Thinking Machines positions Inkling as a base to customise rather than a general-purpose model to rent, and its stated argument is that AI an organisation can adapt for itself will outperform one-size-fits-all models on that organisation's own work. Conceding the benchmark race is what makes that claim coherent.

Does using a model built on Chinese architecture create a compliance problem?

The architecture is a published design, not a data or service dependency, and following it does not route your data anywhere. Inkling's own provenance is worth recording in your AI inventory, since the EU AI Act's transparency obligations concern documentation of the systems you deploy. The practical exposure sits in where you run the model and what data you send it, not in whose paper the expert-routing design came from.

Should we fine-tune an open model instead of paying for a hosted one?

Only if your use case is narrow, repetitive and high volume, and only if you have proprietary training data, an evaluation set drawn from your real work, and someone accountable for the model in production. With those, ownership can beat metered access and the cost stops rising with usage. Without them, renting a hosted model remains the cheaper and more honest choice.

Inkling will not top a leaderboard, and Thinking Machines said so before anyone else could. What it offers instead is a base you can take, shape and keep, built openly on work published by labs on the other side of an export-control debate. The lesson is not that open weights beat hosted models. It is that a vendor willing to tell you what its product is not has given you something to plan with, while a vendor selling you this month's benchmark has given you a number that expires. Decide which of those you are actually buying, and the model choice mostly makes itself.

Thinking Machines LabMira MuratiInklingOpen weightsFine-tuningDeepSeekModel provenanceTinker

DeepSeek Charges Double While Europe Works

DeepSeek V4 launches in mid-July with peak-hour pricing that doubles token rates from 14:00 to 18:00 Beijing time. That window is 08:00 to 12:00 in Frankfurt and 07:00 to 11:00 in London, so European teams pay the surcharge during their own working morning. The deepseek-chat and deepseek-reasoner names retire on 24 July.

2 min read

Artificial Intelligence

AI Tokens Now Have a Rush Hour

DeepSeek will release V4 in mid July with the first time-based AI API pricing: rates double during Beijing business hours. Why AI tokens are becoming a utility, and how European buyers gain a clock advantage.

3 min read

Digital Sovereignty

The EU Council Revived the Rule That Scans Private Messages

The EU Council revived the lapsed rule that lets providers scan private messages with AI. For owners, confidential communication is now a governance choice.