In This Article
Key Takeaways
- Thinking Machines Lab released Inkling on July 15, 2026 — its first in-house foundation model, and its first open-weights release.
- The model card lists 975B total parameters with 41B active per token, a 66-layer mixture-of-experts transformer, pretraining on 45 trillion tokens, and a context window of up to 1M tokens.
- The weights are Apache 2.0 — a permissive licence permitting commercial use, modification and redistribution — on a model near the trillion-parameter mark.
- Artificial Analysis scored Inkling 41 on its Intelligence Index at release. The company itself says the model “is not the strongest overall model available today, open or closed.”
Thinking Machines Lab published its first in-house foundation model on Wednesday, July 15, 2026. The model is called Inkling, its model card lists “975B total parameters, 41B active,” and the weights ship under the Apache 2.0 licence. That last detail is the news. Permissive licensing on a small or mid-sized open model is routine; permissive licensing on a model this large is not.
One qualifier belongs up front, because coverage blurred it. Inkling is the lab's first foundation model, not its first product. Thinking Machines — founded in February 2025 by former OpenAI CTO Mira Murati, per its public record — had already shipped Tinker, a fine-tuning API, and Tinker is how it earns revenue.
The specifications
The following comes from the company's own model card. Inkling is a 66-layer decoder-only transformer with a sparse mixture-of-experts backbone, routing each token to 6 of 256 experts plus 2 shared experts. It was pretrained on 45 trillion tokens spanning text, images, audio and video, per the announcement post. It accepts text, images and audio; it emits text only. The downloadable weights support up to 1M tokens of context, while Artificial Analysis reports 256K on the hosted Tinker API.
Size has a hardware consequence. The Register reported that the model “requires more than two terabytes of GPU memory — a quantity present in around eight of Nvidia's B300 accelerators, or sixteen H200s,” and that the NVFP4 quantised build runs on half that. We found that estimate in one outlet only, so treat it as a single-source figure rather than a vendor spec.
How it scores
The one independent number available at release came from Artificial Analysis, which placed Inkling at 41 on its Intelligence Index and called it “the leading open weights release from a U.S. lab.” For context, the same index scored Nvidia's Nemotron 3 Ultra at 38, Gemma 4 31B at 29, and gpt-oss-120b at 24.
The model card itself reports AIME 2026 at 97.1%, GPQA Diamond at 87.2%, SWE-bench Verified at 77.6%, MMMU Pro at 73.5%, VoiceBench at 91.4%, SimpleQA Verified at 43.9%, Terminal Bench 2.1 at 63.8%, and Humanity's Last Exam with tools at 46.0%. Those are self-reported and not independently replicated at the time of writing — a starting hypothesis for your own evaluation, not a settled ranking.
Thinking Machines is unusually direct about where it sits. Its announcement states: “Inkling is not the strongest overall model available today, open or closed.” Labs rarely write that sentence about their own launch.
The model is not the product
TechCrunch framed the release as a bet against one-size-fits-all AI: the company markets Inkling “less as a finished product than as a starting point, something for organizations to fine-tune themselves through Tinker.” Giving the weights away and charging for customisation is a coherent business, not a contradiction. TechCrunch cites hedge fund Bridgewater Associates as one such user, reporting 84.7% on financial reasoning tests at roughly a fourteenth of the cost — figures the outlet notes were the companies' own evaluations.
On funding, one number is solid: roughly $2 billion at a $12 billion valuation, closed in July 2025, with Nvidia, AMD, Cisco and Jane Street among the investors. A much larger follow-on has been widely reported, but TechCrunch says it had stalled by January without closing, so we are not treating it as the company's valuation.
Not everyone found the release fully transparent. Developer Simon Willison wrote that the model card is “much shorter than I've come to expect from US AI labs” and that its training-data section has “almost nothing of interest in it.” The lab acknowledges training on material subject to intellectual property protection without detailing the datasets.
Update: Inkling-Small, two weeks later
Added July 30, 2026
Thinking Machines released a second model, Inkling-Small, on July 30, 2026, per its model card. It lists 276B total parameters and 12B active across a 42-layer mixture-of-experts transformer, the same 1M-token context, the same text/image/audio input, and the same Apache 2.0 licence. Artificial Analysis scored it 40 — one point below the 975B flagship at roughly a quarter of the size. Some coverage published on July 31 dated the release to that day; the model card says July 30.
The two Inkling models, as documented
| Inkling | Inkling-Small | |
|---|---|---|
| Released | July 15, 2026 | July 30, 2026 |
| Parameters | 975B total / 41B active | 276B total / 12B active |
| Layers | 66 | 42 |
| Context | Up to 1M tokens | Up to 1M tokens |
| Licence | Apache 2.0 | Apache 2.0 |
| AA Intelligence Index | 41 | 40 |
Why it matters
What follows is our analysis rather than reported fact. Three things stand out for practitioners.
The licence is the story, not the score. A model that ranks fourth or fifth on a leaderboard but carries Apache 2.0 terms reaches work the top of the leaderboard cannot: on-premises deployment, air-gapped environments, regulated data that cannot leave a boundary, and products whose economics break at frontier API prices. When deployment is the constraint, licence terms outrank a few index points. Our guide to open-weight versus frontier API works through that trade.
Inkling-Small is the more interesting release for most teams. One index point below the flagship at roughly a quarter of the parameters changes what hardware you must own. Few organisations will stand up eight B300s; a 276B sparse model is a different conversation. A big model to set the ceiling and a small one to make it deployable is now a standard pattern — and it is usually the small one that ships.
Fine-tune-first has a cost you should price. Positioning a base model as raw material assumes you have evaluation data, a training budget and someone to own the pipeline. That beats prompt-engineering a general model when your domain is narrow and your data is proprietary; it is the wrong path when you have neither the data nor the staff. Decide which you are before the licence tempts you into a project you cannot maintain — our model tier guide is a place to start.
Open weights or hosted API?
The choice turns on deployment constraints, data sensitivity and total cost — not on leaderboard position. Our guide lays out where each option genuinely wins.
Read the comparisonSources: Thinking Machines — Inkling model card; Thinking Machines — Introducing Inkling; Thinking Machines — Inkling-Small model card; TechCrunch; Artificial Analysis; The Register; Simon Willison; gHacks; Thinking Machines Lab — Wikipedia. Analysis and framing by Precision AI Academy.
Common questions
When exactly was Inkling released? July 15, 2026, per the model card and the company's announcement post, both dated that day, with same-day independent coverage.
Can I use it commercially? The weights are Apache 2.0, which permits commercial use, modification and redistribution. Read the licence yourself before you build on it — that is true of any model.
Where do I get the weights? Hugging Face carries standard and NVFP4 quantised checkpoints for both models, and both are served on Tinker for fine-tuning.
What does the hosted version cost? Artificial Analysis lists Tinker API pricing of $1.87 per million input tokens, $0.374 cached and $4.68 output at 64K context, and $3.74 / $0.748 / $9.36 at 256K. Those figures come from Artificial Analysis, not from an official price sheet we were able to verify.
Should I switch to it? Only if your constraint is deployment or licensing. If you are happy on a hosted API and your data can live there, an index score of 41 is not by itself a reason to migrate.