Inkling: Thinking Machines Lab's open-weight model

Updated October 25, 2026. An earlier version of this article presented Inkling as a DeepSeek model and attributed unsourced scores to it. That was inaccurate: Inkling was developed by Thinking Machines Lab. The text below has been fully rewritten from the official model card.

Inkling in short

Inkling is the first open-weight model from Thinking Machines Lab, the American company led by Mira Murati. It was released on July 15, 2026, under the Apache 2.0 license, with weights available on Hugging Face, according to its official model card. It is not a DeepSeek model. Some press coverage compared its design to Chinese models such as DeepSeek or Kimi, but the model card describes no technical link: DeepSeek V4 Pro and Kimi K2.5/K2.6 appear there only as comparison models.

Key specifications

• Size: a Mixture-of-Experts transformer with 975 billion total parameters, 41 billion of them active. Each token is routed to 6 of 256 experts, plus 2 shared experts.
• Context: up to one million tokens with the open weights. Artificial Analysis notes a 256,000-token window through Thinking Machines' Tinker API.
• Modalities: text, image and audio inputs; text output only.
• Training: 45 trillion tokens of text, images, audio and video, as noted by Simon Willison, who also points out that the training data documentation is sparse.
• Access: Tinker API, third-party inference providers, or self-hosting with SGLang, vLLM, TokenSpeed, Unsloth or Hugging Face.

Performance

The scores published by Thinking Machines include 77.6% on SWE-bench Verified, 87.2% on GPQA Diamond and 97.1% on AIME 2026. These are self-reported figures, and the comparison table on the model card shows Inkling behind several closed models, such as Claude Fable 5 and GPT-5.6 Sol, on many benchmarks. Thinking Machines itself states that Inkling is not the strongest model available, open or closed, and positions it as a base for fine-tuning. On the independent Artificial Analysis Intelligence Index, it scored 41 at launch, which made it the leading open-weight model from a US lab, ahead of Nemotron 3 Ultra (38).

What open weights under Apache 2.0 change

The Apache 2.0 license allows commercial use, modification and redistribution without revenue thresholds, unlike some competing licenses such as the one for Kimi K3. A company can therefore fine-tune Inkling on its own data and host it on its own infrastructure, which helps keep sensitive data in-house. The constraint is hardware: the model card states at least 2 TB of aggregate GPU memory for the BF16 version (8 NVIDIA B300 or 16 H200), and at least 600 GB for the NVFP4 version. Self-hosting is therefore a matter for technical teams with significant infrastructure, not for a marketing team on its own.

Use cases and limits for a marketing team

For a marketing team, Inkling is mainly of interest through a provider or an internal tool built by the technical teams. Possible uses include:

• A fine-tuned model on the brand's tone: the license allows a model to be adapted to the company's editorial guidelines and content.
• Multimodal analysis: audio and image inputs make it possible to work on podcasts, videos converted to audio or visuals, with a text output.
• Processing confidential data: on internal infrastructure, documents do not leave the company, a point detailed in our guide on cloud AI vs local AI.

The model card lists clear limits: hallucinations of plausible but incorrect content, occasional failures to follow instructions, degraded performance in long conversations and possible cultural or linguistic biases, with uneven performance across languages. It advises against medical, legal or safety-critical use without validation and human oversight, and recommends input and output filters rather than relying on the model's refusals alone. These recommendations fit naturally into an AI governance framework.

Inkling in the open-weight market

Inkling arrived a day before Moonshot AI unveiled Kimi K3, at a time when the most capable open-weight models came mostly from Chinese labs such as DeepSeek, Moonshot and Z.ai. To understand DeepSeek's positioning, see our OpenAI vs DeepSeek comparison. Thinking Machines also announced a smaller version, Inkling Small (276 billion parameters, 12 billion active), according to Simon Willison.

About Brandeploy

Whatever model you use to produce text, the visuals that go out to your channels still have to respect your brand. Brandeploy is a creative automation platform that helps marketing teams produce and adapt their assets, from banners to localized versions, while keeping a consistent brand identity. Book a demo to see the Brandeploy platform in action.

FAQ

Who developed the Inkling AI model?

Inkling was developed by Thinking Machines Lab, the American company led by Mira Murati. It is not a DeepSeek model. Released on July 15, 2026, it is the lab's first open-weight model.

Is Inkling open source?

Inkling's weights are published on Hugging Face under the Apache 2.0 license, which allows commercial use, modification and redistribution without revenue thresholds. The training data documentation remains sparse.

What are Inkling's main specifications?

According to its model card, Inkling is a Mixture-of-Experts transformer with 975 billion total parameters, 41 billion of them active, and a context of up to one million tokens with the open weights. It accepts text, image and audio inputs and produces text only.

How does Inkling perform?

Thinking Machines reports 77.6% on SWE-bench Verified and 87.2% on GPQA Diamond. The lab acknowledges that Inkling is not the strongest model available and positions it as a base for fine-tuning. Artificial Analysis gave it a score of 41 on its Intelligence Index at launch, the best result for an open-weight model from a US lab at the time.

What are Inkling's limits?

The model card mentions hallucinations, occasional failures to follow instructions, degraded performance in long conversations and possible cultural or linguistic biases. Self-hosting requires at least 600 GB of GPU memory in NVFP4 and 2 TB in BF16. Human oversight and input and output filters are recommended.