The release of Inkling, a general-purpose artificial intelligence model developed by Thinking Machines Lab, marks a milestone in the open-weight AI ecosystem. Founded by former OpenAI CTO Mira Murati, this San Francisco-based startup has unveiled a model that not only competes in performance but also offers a US-based alternative in a market dominated by Chinese developments. With 975 billion total parameters —41 billion of which are active during inference— and a context window of up to one million tokens, Inkling is designed for multimodal tasks integrating text, images, audio, and video.
The mixture-of-experts (MoE) architecture allows Inkling to optimize computational resource usage by activating only a fraction of its parameters per processing step. Combined with a reasoning-effort setting ranging from 0.2 to 0.99, developers gain fine-grained control over the balance between performance and generated token count. In internal tests, the company demonstrated that Inkling matches Nvidia Nemotron 3 Ultra on Terminal Bench 2.1 while generating approximately one-third fewer tokens. Moreover, its 77.6% score on SWE-Bench Verified —though behind DeepSeek V4 Pro and GLM 5.2— surpasses Nemotron, establishing it as a solid option for software engineering tasks.
For enterprises, the full open weights of Inkling on platforms like Hugging Face —both as the original checkpoint and as a quantized NVFP4 checkpoint— represent an opportunity to deploy AI on private infrastructure, meeting data privacy and sovereignty requirements. However, the implementation cost is not trivial: running the full model requires a GPU cluster with at least 2 TB of aggregated VRAM, which translates to eight Nvidia B300 GPUs or 16 H200 GPUs. The quantized version lowers the requirement to 600 GB, enabling execution on four B300 or eight H200 GPUs. For many organizations, this infrastructure cost may make closed-model APIs more economical, though the flexibility of an open-weight model is hard to match.
Thinking Machines has also previewed Inkling-Small, with 276 billion total parameters and 12 billion active, a lighter version that promises to reduce latency and infrastructure costs without sacrificing useful performance on key workloads. The company plans to release its full weights after final testing. This smaller model could be an entry point for many companies seeking to integrate advanced AI without major hardware investments.
From a safety and governance perspective, Inkling was trained for calibration, instruction following, and resistance to censorship. It scored 98.6% on StrongREJECT, a test measuring whether the model refuses clearly harmful requests. However, analysts like Pareekh Jain warn that fine-tuning can weaken safety filters, so companies must retest safety after customizing the model. The ability to self-host and modify the model means that customized versions can diverge from the official one over time without receiving automatic updates. Therefore, CIOs must ensure that every AI agent action is logged, auditable, and subject to human approval for high-risk tasks.
This is where the expertise of companies like Q2BSTUDIO comes into play. Specializing in custom software development and AI integration, our team can help organizations deploy models like Inkling on cloud environments such as AWS or Azure, ensuring the infrastructure meets scalability and security requirements. In addition, we offer cybersecurity services to audit and strengthen customized models, and Business Intelligence solutions with Power BI to extract maximum value from data processed by AI agents.
The launch of Inkling not only broadens the range of options in the open-weight model market but also reinforces the trend toward enterprise adoption of AI with full data control. At Q2BSTUDIO, we understand that every organization has unique needs, so we provide consulting to select the most suitable model and infrastructure, whether Inkling, GPT, Claude, or other alternatives. Combining an open-weight model with an expert technology partner enables companies to innovate quickly while maintaining security and regulatory compliance.
In summary, Inkling from Thinking Machines Lab represents a significant step forward in democratizing high-performance AI. Its open nature, multimodal capability, and reasoning control make it a powerful tool for enterprise applications such as knowledge-intensive copilots, multimodal customer service, workflow automation, and agentic tasks. However, successful implementation depends on a solid strategy considering infrastructure costs, security, and governance. With support from companies like Q2BSTUDIO, organizations can overcome these challenges and fully leverage the potential of open-weight AI.
