Article · 2026-07-16 · 6 moments

Inkling: Our open-weights model

Inkling: Our open-weights model Mira Murati's Thinking Machines Lab just released their first open-weights model. Inkling is "a Mixture-of-Experts transformer with 975B total parameters, 41B active" - an Apache-2.0 licensed multimodal model trained on 45 trillion tokens of text, images, audio and video. They're also promising Inkling-Small, a 276B (12B active) model, but that's still being tested and the weights will be released "once that work is complete". The model card is much shorter tha ✦ AI generated

Read original ↗

GUEST Thinking Machines Lab GUEST Inkling HOST Simon Willison

Data◆

Inkling is a Mixture-of-Experts transformer with 975B total parameters and 41B active, an Apache-2.0 licensed multimodal model trained on 45 trillion tokens of text, images, audio and video.

Thinking Machines Lab released Inkling, a large Apache-2.0 licensed open-weights MoE multimodal model.

transcript

Simon Willison: Inkling is "a Mixture-of-Experts transformer with 975B total parameters, 41B active" - an Apache-2.0 licensed multimodal model trained on 45 trillion tokens of text, images, audio and video.

Claim◆

Inkling is a welcome, competitive addition to the US open-weights ecosystem, being Apache-2.0 licensed and comparable to open-weight models coming out of China, joining NVIDIA Nemotron and Gemma 4 as viable contenders.

Simon Willison praises the release as competitive with Chinese open-weight models and a good new option alongside Nemotron and Gemma 4.

transcript

Simon Willison: There's a lot to like about this release. It's Apache-2.0 licensed, and looks competitive with the open weight models coming out of China - it's good to see the US open weights ecosystem gain a new viable contender to join NVIDIA Nemotron and Gemma 4.

provides context · 1supports · 1

Claim◆

Inkling is not the strongest overall model available today, open or closed; instead its multimodal capabilities, efficient thinking, and availability on Tinker make it a good open-weights base for customization via fine-tuning.

Thinking Machines Lab frames Inkling not as a state-of-the-art model but as a strong, efficient base meant to be fine-tuned through their Tinker platform.

transcript

Thinking Machines Lab: Inkling is not the strongest overall model available today, open or closed. Instead, a combination of qualities makes it a good open-weights base for customization: multimodal capabilities, efficient thinking, and availability on Tinker for fine-tuning.

Fact

Thinking Machines Lab's training data documentation only vaguely states that its datasets include public domain and possibly IP-protected content gathered from the open internet, public repositories, and third parties.

Simon Willison observes that Inkling's model card and training data documentation are unusually thin, offering only generic boilerplate about data sourcing.

transcript

Thinking Machines Lab: Thinking Machines Lab's services were developed using publicly available content obtained from the open internet and publicly accessible data repositories. Certain datasets were also obtained from third parties.

Fact

Thinking Machines Lab will release Inkling-Small, a 276B (12B active) model, once testing on it is complete.

A smaller companion model, Inkling-Small, is promised but withheld pending further testing.

transcript

Simon Willison: They're also promising Inkling-Small, a 276B (12B active) model, but that's still being tested and the weights will be released "once that work is complete".

Anecdote

When shown a JPEG rendering of the SVG pelican-on-a-bicycle image it had generated, Inkling misidentified the bird as a 'stork or seagull' rather than recognizing it as a pelican.

Testing Inkling's multimodal self-description, Simon Willison found the model failed to recognize its own SVG pelican drawing, calling it a stork or seagull instead.

transcript

Inkling: This is a cheerful, flat-vector cartoon illustration featuring a white bird riding a bicycle across a green landscape. The Bird: The central figure is a stylized white bird—resembling a stork or seagull—with a plump, rounded body.

Highlight slides

Inkling: Thinking Machines Lab's New MoE Model✦ from: Inkling is a Mixture-of-Experts transformer with 975B total parameters and 41B active, an Apache-2.0 licensed multimodal model trained on 45 trillion tokens of text, images, audio and video.Inkling: Not the Strongest, But Built to Customize✦ from: Inkling is not the strongest overall model available today, open or closed; instead its multimodal capabilities, efficient thinking, and availability on Tinker make it a good open-weights base for customization via fine-tuning.Inkling: A New US Open-Weights Contender✦ from: Inkling is a welcome, competitive addition to the US open-weights ecosystem, being Apache-2.0 licensed and comparable to open-weight models coming out of China, joining NVIDIA Nemotron and Gemma 4 as viable contenders.Open-Weights, Multimodal by Design✦ from: Inkling is a Mixture-of-Experts transformer with 975B total parameters and 41B active, an Apache-2.0 licensed multimodal model trained on 45 trillion tokens of text, images, audio and video.Why Inkling Works as a Base Model✦ from: Inkling is not the strongest overall model available today, open or closed; instead its multimodal capabilities, efficient thinking, and availability on Tinker make it a good open-weights base for customization via fine-tuning.Joining the Open-Weights Field✦ from: Inkling is a welcome, competitive addition to the US open-weights ecosystem, being Apache-2.0 licensed and comparable to open-weight models coming out of China, joining NVIDIA Nemotron and Gemma 4 as viable contenders.

Related episodes

Article · Interconnects

Inkling is a Mixture-of-Experts transformer with 975B total parameters and 41B active, an Apache-2.0 licensed multimodal model trained on 45 trillion tokens of text, images, audio and video.

Inkling is a welcome, competitive addition to the US open-weights ecosystem, being Apache-2.0 licensed and comparable to open-weight models coming out of China, joining NVIDIA Nemotron and Gemma 4 as viable contenders.

Inkling is not the strongest overall model available today, open or closed; instead its multimodal capabilities, efficient thinking, and availability on Tinker make it a good open-weights base for customization via fine-tuning.

Thinking Machines Lab's training data documentation only vaguely states that its datasets include public domain and possibly IP-protected content gathered from the open internet, public repositories, and third parties.

Thinking Machines Lab will release Inkling-Small, a 276B (12B active) model, once testing on it is complete.

When shown a JPEG rendering of the SVG pelican-on-a-bicycle image it had generated, Inkling misidentified the bird as a 'stork or seagull' rather than recognizing it as a pelican.

Kimi K3: The open-weights escalation

Community perspectives on Elastic vs AWS (Interview)

Fable's Back, AI Engineer Recap, & SambaNova