Thinking Machines выпустила Inkling — открытую MoE-модель на 975 млрд параметров

Mira Murati's startup Thinking Machines Lab released its first open-weights model, Inkling, on July 16, 2026 — a Mixture-of-Experts transformer with 975 billion parameters (41 billion active) under the Apache 2.0 license, trained on 45 trillion tokens of text, images, audio, and video.

What is Inkling

Inkling is a multimodal MoE model with 975 billion parameters, of which 41 billion are active on each request. The developer, Thinking Machines Lab, released the weights under the permissive Apache 2.0 license and accompanied the release with an unusually short model card — far more concise than what American AI labs typically publish.

• Developer — Thinking Machines Lab, the startup of Mira Murati, ex-CTO of OpenAI
• Architecture — Mixture-of-Experts, 975 billion parameters total, 41 billion active
• Data — 45 trillion tokens of text, images, audio, and video
• License — Apache 2.0, fully open weights
• A smaller version, Inkling-Small, has been promised — 276 billion parameters (12 billion active), with weights to follow after testing

Not a flagship, but a base for fine-tuning

Thinking Machines Lab openly acknowledges that Inkling is not the strongest model on the market, either among open or closed models. Instead of racing for the top spot, the team positions it as a convenient base for customization and fine-tuning on its own Tinker platform.

"Inkling is not the strongest model available today, open or closed.

What makes it a good open base for customization is a combination of qualities: multimodality, efficient 'thinking,' and availability in Tinker for fine-tuning," the Thinking Machines Lab model card states.

Documentation on the training data is nearly empty: the company only notes that it used content from the open internet and publicly available repositories, that some data was obtained from third parties, and that this may include material protected by intellectual property rights.

Why this matters for the open-model market

Inkling strengthens the American open-weights ecosystem, where Chinese releases have recently set the tone. According to developer Simon Willison, who tested the model via the API, Inkling looks competitive against open models from China and joins the few viable American contenders — NVIDIA Nemotron and Gemma 4.

Its multimodality has been confirmed in practice: Willison generated an SVG image via the API, then asked the same model to describe it — Inkling broke down the image in detail, mistaking the drawn bird for a stork or a seagull.

What this means

An open multimodal base from a team of former OpenAI executives lowers the barrier to entry for those who want to build and fine-tune their own models without depending on closed APIs. Thinking Machines' bet is not on benchmarks, but on the convenience of customization through Tinker.

Frequently asked questions

What is Inkling?

Inkling is Thinking Machines Lab's first open-weights model, a multimodal Mixture-of-Experts transformer with 975 billion parameters (41 billion active), trained on 45 trillion tokens and released on July 16, 2026.

How many parameters does the Inkling model have?

975 billion parameters in total, of which 41 billion are active on each request. The promised smaller version, Inkling-Small, has 276 billion parameters (12 billion active).

Can Inkling be used for free?

Yes. The weights are open under the Apache 2.0 license — the model can be freely downloaded, used, and fine-tuned, including on Thinking Machines Lab's Tinker platform.