Thinking Machines Open Sources First Multimodal Language Model, Inkling
New AI model focused on low cost and censorship resistance makes waves in the industry.
1A New Open-Source Revolution
Thinking Machines has open-sourced its multimodal language model, Inkling, under the Apache 2.0 license. This move aims to meet the demand for customizable and on-premise AI models. The model stands out as a Mixture-of-Experts (MoE) system capable of processing text, image, and audio data.
2Performance and Competition
Inkling demonstrates high performance in third-party benchmarks, particularly in software engineering and voice understanding. However, competitors like China's leading AI labs outperform Inkling in certain areas, showcasing the fierce competition in the open-weight model landscape.
3Resistance to Censorship
Inkling distinguishes itself by providing direct answers on topics subject to censorship. This feature offers businesses a reliable option for accurate information, even on controversial or sensitive subjects, making it attractive for organizations prioritizing data security and transparency.
4Multimodal Capabilities
Inkling is designed as a native multimodal model, processing text, image, and audio data. These capabilities enable its use across a wide range of applications. It performs competitively in visual and auditory data processing compared to rivals.
5Cost Control
Thinking Machines designed Inkling with a focus on balancing cost and performance. Unlike traditional black-box scaling strategies, the model offers a programmable thinking effort mechanism, allowing users to manage costs effectively.
6Inkling-Small
Thinking Machines also previewed Inkling-Small, a lighter model optimized for workloads where low latency and cost are critical. With 276 billion parameters, this alternative is ideal for smaller-scale applications.
7Benchmarking Inkling
Inkling shows varied results in benchmark tests. For example, it outperforms DeepSeek V4 Pro in mathematical problems but falls short in coding and text-based reasoning against GLM 5.2. Its multimodal capabilities help maintain a balanced performance overall.
8U.S.-Based Competitors
Inkling performs strongly against U.S.-based open-source rivals like Nvidia Nemotron 3, surpassing them in many areas. However, it lags behind closed-source giants such as Claude Fable 5 and GPT 5.6 Sol in certain benchmarks.
9Real-World Applications
Thinking Machines designed Inkling for adaptability and efficiency in real-world workflows. Its standout feature is the programmable thinking effort mechanism, allowing users to dictate how much the AI 'thinks' before responding.
10Future Potential
As an open-source foundation model, Inkling has the potential to drive future innovations. Its multimodal integration and censorship resistance could make it indispensable in the business world. Thinking Machines' approach hints at the future direction of AI.
11Conclusion and Implications
The release of Inkling marks a significant milestone in the AI world. This open-source, multimodal model offers new opportunities for businesses. Thinking Machines' strategy provides insights into how AI might evolve in the coming years.