Thinking Machines Releases Inkling: Open-Source Multimodal AI Focused on Cost and Censorship Resistance

AI startup Thinking Machines has launched Inkling, its first major open-source multimodal language model, under an Apache 2.0 license. Designed for enterprises, Inkling aims to offer customization, control, and on-premises deployment capabilities. The model boasts high performance on benchmarks for software engineering and voice understanding, though it is not always state-of-the-art compared to some rivals. A key differentiator is Inkling's design to provide direct answers on topics that may be subject to censorship, positioning it as a trustworthy option for factual outputs regardless of controversy. Inkling is a 975 billion parameter, natively multimodal Mixture-of-Experts (MoE) system capable of processing text, images, and audio, with weights available on Hugging Face and the company's API. It features a novel "controllable thinking effort" mechanism to balance cost and performance, allowing developers to adjust the AI's reasoning budget. A smaller version, Inkling-Small (276 billion parameters), is also available for low-latency and cost-sensitive workloads. While Inkling performs strongly against U.S. open-weight competitors like Nvidia Nemotron 3, it trails behind leading Chinese models such as GLM 5.2 and DeepSeek V4 Pro in areas like pure coding and complex reasoning. It also falls short of top-tier closed-source models like Claude Fable 5 and GPT 5.6 Sol in peak reasoning and software engineering autonomy, but remains competitive in multimodality. The model's training included an emergent phenomenon called "chain of thought condensation," which compresses reasoning steps for reduced latency. Thinking Machines also emphasized Inkling's epistemics, including its calibration, instruction following, and resistance to censorship, demonstrating strong non-compliance with propaganda and censorship evaluations while maintaining defenses against malicious requests.