Thinky Releases Open Inkling 975B-A41B Model Under Apache 2.0 License

In July 2026, Thinky released Inkling, a multimodal language model with 975 billion parameters and open weights under the Apache 2.0 license. It is the company's first full-fledged large language model release, and it launched with open weights from day one.

What is Inkling?

Inkling uses a Mixture-of-Experts architecture: of its 975 billion parameters, only 41 billion are active on each request (hence the designation 975B-A41B). An MoE model routes each token to a subset of "experts," so the model is flagship-sized while inference costs resemble those of a much smaller dense model. For the user, this means Inkling answers like a 975-billion-parameter system but is closer to a 41-billion-parameter model in speed and per-request cost. That balance is why the largest labs use MoE. Inkling is multimodal, so it works with more than just text.

Alongside the flagship, Thinky also released a lighter version, Inkling-Small, with 276 billion parameters and 12 billion active (276B-A12B). Both models use the same MoE approach and differ in size; the smaller one is aimed at people who lack the hardware for the larger one.

• Flagship: Inkling, 975B total / 41B active (975B-A41B), multimodal
• Compact version: Inkling-Small, 276B / 12B active (276B-A12B)
• License: Apache 2.0 — open weights, commercial use without restrictions
• Positioning: billed as the best American open model under Apache 2.0
• Developer: Thinky, for which this is the first full-fledged LLM release

Why open weights matter

Open weights mean anyone can download the model, run it on their own hardware, fine-tune it, and build it into products without going through a vendor's API. The Apache 2.0 license is one of the most permissive: it allows commercial use, modification, and distribution with practically no restrictions, unlike the stricter "community" licenses some competitors use. For startups and large companies alike, this removes the legal questions that keep some open models out of production.

According to the AINews newsletter from Latent Space, the release turned out strong.

"Thinky's first full-fledged LLM release is a bombshell, and open weights come as a bonus," the AINews newsletter from

Latent Space says.

Where it fits among open models

Inkling lays claim to the title of best American open model under the Apache 2.0 license, which is how the source describes it. Over the past two years, Chinese labs, chiefly DeepSeek, Qwen, and Kimi, have mostly set the bar in open weights. A strong American model with 975 billion parameters and a fully open Apache license is a notable shift in that landscape.

The key point is not only the open weights themselves but the license. Many popular open models ship under their own "community" licenses with limits on user counts or use cases. Apache 2.0 contains no such caveats, so an American model released specifically under this license removes some of the legal risk for businesses.

The pair of larger and smaller models (975B-A41B and 276B-A12B) covers different scenarios: the flagship for maximum quality, Inkling-Small for running on more modest hardware.

What this means

An open multimodal model of this size from an American team gives developers another strong option that they can deploy locally and use commercially without license caveats. Having two sizes lets teams pick a model to fit their compute budget, and for the open-weights market it means more competition from outside Chinese labs.

Frequently asked questions

What is Inkling 975B-A41B?

It is Thinky's flagship multimodal model built on a Mixture-of-Experts architecture: 975 billion parameters in total and 41 billion active per request. The weights are open under the Apache 2.0 license.

Is Inkling free?

Yes. The weights are published openly under Apache 2.0, so you can download the model, run it, and use it in commercial products for free. You only pay for your own compute hardware or cloud.

How does Inkling-Small differ from the larger model?

Inkling-Small is smaller: 276 billion parameters versus 975 billion, and 12 billion active versus 41 billion. It is designed to run on less powerful hardware, while the larger model targets maximum quality.