TL;DR
Thinking Machines has publicly released its latest foundation model, Inkling, a 975-billion-parameter multimodal transformer. The model is openly available on Hugging Face under Apache 2.0, but its training data and policies raise questions about true openness. This development highlights ongoing debates over open AI models’ accessibility and restrictions.
Thinking Machines has officially released its first foundation model, Inkling, under an open-source license. The model, a 975-billion-parameter multimodal transformer, is now available on Hugging Face with full weights, marking a significant step in open AI development. This move is notable because it directly challenges the industry norm of proprietary models and raises important questions about access, licensing, and use restrictions.
Inkling is a Mixture-of-Experts transformer with 975 billion total parameters and 41 billion active. It supports a one-million-token context window and was trained on 45 trillion tokens across text, images, audio, and video modalities. Unlike many models that bolt vision capabilities onto language models, Inkling’s multimodal components were trained from scratch, allowing for native multimodal input—text, images, and audio—processed jointly without encoder-based adapters.
The model’s weights are released under Apache 2.0 license, enabling download, modification, and commercial use. However, sources indicate that Thinking Machines maintains a separate Model Acceptable Use Policy that restricts surveillance, deception, and automated decision-making affecting individuals, creating a layer of restrictions beyond the open license. The model was trained with a hybrid optimizer on NVIDIA systems, and its performance was measured using external benchmarks, including a 97.1% score on AIME 2026 and 87.2% on GPQA Diamond.
While the open weights are available, the training data and full pipeline are not published, and some reports suggest restrictions through a separate policy, which complicates the notion of “open source.” The release also included a smaller variant, Inkling-Small, which matches or surpasses the larger model on several benchmarks, with full weights expected after testing.
Implications of Open Release and Licensing Restrictions
This release signifies a notable shift in AI development philosophy by openly sharing large-scale models while maintaining certain restrictions through separate policies. It raises questions about what truly constitutes open source in AI, especially when restrictions on use are layered on top of open licenses. For developers and organizations, this model offers greater control and flexibility, enabling local deployment and modification, which is critical for sensitive domains like public safety or geospatial analysis. However, the potential restrictions in the Model Acceptable Use Policy could limit some applications, making it essential for users to scrutinize the licensing and policies before deployment.
The move also pressures other AI companies to clarify their licensing and openness strategies, especially amid ongoing debates over model safety, misuse, and transparency. As the industry grapples with balancing openness and control, Inkling’s release exemplifies a pragmatic approach—sharing powerful models while attempting to regulate their use through layered policies.
As an affiliate, we earn on qualifying purchases.
Background on Open Models and Industry Norms
Over the past year, the AI industry has seen increasing tension between proprietary development and open-source sharing. Major players like OpenAI and Meta have largely kept their most powerful models closed or restricted, citing safety and misuse concerns. In contrast, some organizations have pushed for open weights to foster innovation, transparency, and local deployment. Thinking Machines, founded by former OpenAI CTO, has taken a different approach by releasing Inkling’s weights openly on Hugging Face, under Apache 2.0, but with indications of layered restrictions through a separate policy.
This approach responds to industry debates about the true meaning of open source in AI, especially when models are accompanied by policies that limit certain types of use. The recent release follows a pattern of cautious openness, balancing transparency with control, and reflects broader trends toward more nuanced licensing frameworks in the AI community.
Prior to this, most large models were either proprietary or released with limited access. Inkling’s openness, combined with its extensive multimodal capabilities and performance benchmarks, marks a notable milestone in this evolving landscape.
“Our goal was to provide a powerful, open model that developers can freely adapt, while ensuring responsible use through our policies.”
— Thinking Machines spokesperson
As an affiliate, we earn on qualifying purchases.
Remaining Questions About Inkling’s Use and Restrictions
It is still unclear how the Model Acceptable Use Policy will be enforced and whether it will significantly restrict certain applications. The full training data and pipeline are not published, raising questions about transparency and reproducibility. Additionally, the actual scope of restrictions, especially in sensitive domains like surveillance or automated decision-making, remains to be verified through the policy documentation.
Further clarity is needed on how the layered restrictions interact with the open license, and whether users can freely modify and deploy the model without risking violations of the policy.
large language model training kits
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Next Steps for Developers and Industry Watchers
Expect independent evaluations of Inkling’s performance across diverse benchmarks and real-world applications. The full weights of Inkling-Small are anticipated soon, which will provide additional insights into the model’s capabilities. Users should scrutinize the Model Acceptable Use Policy before deploying the model in sensitive contexts.
Industry observers will monitor how other organizations respond—whether they adopt similar layered licensing or push for fully open models. Additionally, regulatory discussions around AI transparency and responsible use are likely to intensify as models like Inkling demonstrate new possibilities and challenges.
Domain-Specific Small Language Models: Efficient AI for local deployment
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What makes Inkling different from other large language models?
Inkling is a 975-billion-parameter multimodal transformer released under Apache 2.0, with native support for text, images, and audio, trained from scratch for multimodal input, and available for local deployment.
Is Inkling truly open source?
While the weights are released under Apache 2.0, the training data and full pipeline are not published, and a separate use policy may restrict certain applications, complicating the notion of full openness.
What restrictions might apply to Inkling’s use?
Reports suggest a Model Acceptable Use Policy that prohibits surveillance, deception, and automated decisions affecting individuals, but the details are not fully confirmed; users should review the policy before use.
How does this release impact the AI industry?
It sets a precedent for layered licensing—combining open weights with restrictions—potentially influencing how future models are shared and governed, especially in sensitive or regulated domains.
Source: ThorstenMeyerAI.com
