Inkling Explained: Thinking Machines' First Open-Weights Model and Why It Aims for Customization Over the Crown

On July 15, 2026, Thinking Machines Lab released Inkling, its first model, as open weights under the Apache license. The framing that accompanied it was unusual. Rather than claiming a new state of the art, the company said plainly that Inkling is not the strongest model available today, open or closed. Instead it pitched a well-rounded, multimodal, efficient model designed to be customised and owned rather than merely rented.

That combination, a serious open-weights release from a heavily funded lab that deliberately does not chase the leaderboard crown, makes Inkling one of the more interesting model launches of the year. This piece breaks down what it actually is, how it performs by independent measurement, and why the strategy is as much the point as the specifications.

What Inkling is, technically

Inkling is a mixture-of-experts (MoE) transformer with 975 billion total parameters but only 41 billion active on any given token. That design is the crux of modern efficient models: total parameters set the ceiling on knowledge and capability, while active parameters set the compute cost per token. A large-but-sparse model aims to be knowledgeable without being as expensive to run as its raw size suggests. Under the hood, Thinking Machines uses 256 routed experts plus two shared experts per layer with six routed experts active per token, relative positional embeddings rather than the now-common RoPE, and short convolutions inside the attention layers.

It was pretrained on 45 trillion tokens spanning text, images, audio and video. In practice it takes text, image and audio as input and produces text output, including code and structured data. Context length is up to 1 million tokens in the open-weights release, and 256K through the company's hosted Tinker API. A smaller companion, Inkling-Small, a 276-billion-parameter MoE with 12 billion active, shipped as a preview and reportedly matches or exceeds its larger sibling on many benchmarks, a reminder that parameter count and measured performance are not the same thing.

How it actually performs

Self-reported benchmarks always flatter the vendor, so the more useful reference is the independent evaluation from Artificial Analysis, which measured Inkling shortly after release.

By that independent yardstick, Inkling scores 41 on the Artificial Analysis Intelligence Index, which Artificial Analysis called the leading open-weights release from a U.S. lab. It edges out the previous U.S. open-weights leader, Nvidia's Nemotron 3 Ultra at 38, and sits well ahead of Google's Gemma 4 31B at 29 and OpenAI's gpt-oss-120b at 24. On agentic evaluations it posted an Elo of 1238 on GDPval-AA v2, ahead of strong Chinese open-weights models Kimi K2.6 and DeepSeek v4 Flash, and it edged them on a banking agent benchmark too.

Two points of discipline matter when reading this. First, a three-point lead on a composite index is a lead, not a rout; Inkling is at the front of the U.S. open-weights pack, not on a different tier from it. Second, the framing is specifically about U.S. open-weights models. The strongest open-weights systems overall include Chinese releases, and Inkling's own reported numbers on the hardest tests, such as a 29.7 percent on the deliberately brutal Humanity's Last Exam, confirm the company's own point that this is a well-rounded model rather than a frontier-topping one.

The efficiency angle

Where Inkling's numbers get genuinely interesting is not peak score but cost to reach it. Artificial Analysis found Inkling averaged about 25,000 output tokens per Intelligence Index task, against roughly 37,000 to 43,000 for several strong open-weights peers. Thinking Machines makes a similar claim on agentic coding, reporting comparable results to Nemotron at around a third of the tokens.

This matters because tokens are the unit of both cost and latency. A model that reaches the same answer in fewer tokens is cheaper to run and faster to respond, and in agentic workflows, where a single task can spin through long internal chains of reasoning and tool calls, that efficiency compounds. Inkling also exposes a controllable thinking effort dial, letting a user trade capability against token spend per task, which turns efficiency from a fixed property into a knob. For anyone paying per token, a well-rounded model that is deliberately economical with them is a meaningful proposition, arguably more so than a few extra points on a benchmark.

How it was trained, and why that is part of the story

Thinking Machines was unusually open about method. Inkling was pretrained from scratch on Nvidia GB300 NVL72 systems, then shaped with large-scale asynchronous reinforcement learning scaled past 30 million rollouts across long continuous runs, during which reasoning performance reportedly improved log-linearly. The post-training was bootstrapped with supervised fine-tuning on synthetic data generated by open models, including Moonshot AI's Kimi K2.5, before the large-scale RL took over. The company also noted an emergent effect: through RL, the model's chain-of-thought grew more concise over time on its own, shedding grammatical overhead while staying readable.

One detail is worth flagging honestly, because Thinking Machines flagged it itself: using another lab's open model to bootstrap early post-training data is a common and legitimate practice, but it does mean the first model was not built in complete isolation. The company has said its next model will use fully self-contained post-training. Transparency about that lineage is part of what makes the release readable, and it is the kind of methodological candour that is still not universal among model launches.

The part most launches skip: epistemics

The most distinctive design choice is that Inkling was trained to know what it does not know. Thinking Machines trained it with reinforcement learning against proper scoring rules on a large corpus of resolved real-world questions, which is a formal way of rewarding calibrated confidence, being appropriately unsure rather than confidently wrong. It also used dedicated factuality graders, one of which runs its own agentic web search to verify claims, and reports competitive calibration and hallucination figures on independent accuracy tests.

This is easy to overlook next to headline benchmark numbers and easy to overvalue in the abstract, so it is worth stating plainly. Calibration does not make a model smarter, and these are still early, self-and-independently-reported measures rather than a solved problem. But for real deployments, especially agentic ones where a wrong step propagates, a model that signals uncertainty instead of fabricating can be more useful than a higher-scoring model that does not. It is a design priority aimed at production reliability rather than leaderboard placement, which is consistent with the whole positioning.

The strategy behind the specs

Every choice above serves one strategy. Inkling ships as open weights on Hugging Face, including an NVFP4 checkpoint tuned for efficient inference on Nvidia Blackwell hardware, and it is available for fine-tuning on Thinking Machines' Tinker platform, with a free Inkling Playground for chat and broad third-party distribution across inference providers. The business model is not per-token rent on a proprietary model; it is the ecosystem of customization, fine-tuning and hosting around a model you can take and adapt.

That reframes how to judge Inkling. Against a pure capability question, is this the smartest model, the answer is no, and the company agrees. Against the question it is actually built for, is this a strong, efficient, honest, well-rounded base that an organisation can adapt and own, the independent numbers suggest a credible yes for the U.S. open-weights tier. Those are different questions, and conflating them is the easiest way to misread the release.

For teams deciding what to build on, Inkling widens the menu rather than replacing it. It joins a growing field of capable open-weights models, alongside strong Chinese releases and other U.S. efforts, that increasingly makes hard-wiring a product to a single proprietary model look like a choice rather than a necessity. The practical way to exploit that is to keep the option open. A model-agnostic workspace such as Metir AI brings open-weights models like Inkling together with closed frontier models in one place, so a team can route a customizable, cost-sensitive job to an adaptable model and a hard frontier task to whatever leads today, without committing the whole stack to either.

The bigger picture

Inkling is a clear expression of a specific idea: that for a large share of real-world AI work, the winning attributes are not peak intelligence but adaptability, efficiency, honesty about uncertainty, and ownership. It is a strong, independently validated open-weights model that deliberately declines to compete on the one axis, raw capability, that dominates most headlines.

Whether that bet is right is genuinely unsettled, and depends on how many organisations want to customise and own models rather than rent frozen ones. What Inkling does, cleanly, is give that thesis a concrete, measurable, downloadable form. It is less a challenger for the frontier crown than an argument that the crown is not the only prize worth chasing, and it is now out in the open for anyone to test that argument themselves.

Sources:

• Introducing Inkling | Thinking Machines Lab
• Thinking Machines has released Inkling, the new leading U.S. open-weights model | Artificial Analysis
• Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling | TechCrunch
• Mira Murati's Thinking Machines debuts its first AI model | Axios
• Thinking Machines Launches AI Model Inkling with 975B Parameters | GuruFocus