Thinking Machines Releases Inkling, Its Debut Open-Weights AI Model

Mira Murati's Thinking Machines Lab Enters the Open Arena

Thinking Machines Lab has officially launched Inkling, its inaugural open-weights artificial intelligence model, marking the high-profile startup's first major software release. The model arrives with broad open distribution, with complete model weights published directly on Hugging Face to encourage community adoption and local experimentation.

Founded by former OpenAI Chief Technology Officer Mira Murati, Thinking Machines entered the artificial intelligence landscape backed by massive investor confidence, having raised a 2 billion dollar seed round at a 12 billion dollar valuation with participation from Andreessen Horowitz, Nvidia, AMD, Cisco, and Accel. The release of Inkling represents the company's opening bid to balance frontier research with accessible, customizable developer infrastructure.

Multimodal Agentic Architecture

Inkling is designed from the ground up as a multimodal model capable of processing visual and textual data. A key architectural distinction lies in how the model handles visual perception tasks during inference. Rather than treating visual inputs solely through static token embeddings, Inkling can actively execute Python code to perform programmatic visual operations such as targeted cropping and dynamic zooming.

This dynamic code-execution mechanism bridges standard computer vision with agentic problem-solving. By manipulating images programmatically, the system functions more like a digital analyst using specialized tools than a passive pattern recognition engine. The capability allows developers to build workflows that require fine-grained examination of complex charts, technical schematics, and high-resolution imagery.

The Tinker Platform and Open Inference Support

Alongside the weights release, Thinking Machines introduced Tinker, a dedicated managed fine-tuning platform and training API. Tinker is structured to decouple the engineering overhead of distributed cluster management from model experimentation. Researchers and commercial developers can direct parameter updates and specialized alignment training while Thinking Machines manages the underlying computational scaling.

To ensure widespread immediate deployment across existing hardware stacks, Thinking Machines collaborated with multiple open source inference teams prior to the debut. Inkling includes day-one compatibility with several prominent inference engines:

• vLLM: Supported through collaboration with Inferact for high-throughput serving architectures.
• SGLang: Optimized alongside RadixArk to enable complex structured generation and low-latency agent workflows.
• llama.cpp: Integrated via Unsloth to bring local, efficient quantized execution to everyday desktop hardware.
• TokenSpeed: Tuned with Lightseek for optimized inference throughput.

Shifting Dynamics in Frontier AI Development

The choice to release Inkling under an open-weights paradigm highlights the growing competition among artificial intelligence labs pursuing open developer ecosystems. While closed API frontier providers continue to lock down proprietary model architectures, well-funded alternatives are increasingly releasing functional model parameters to capture mindshare among software engineers and enterprise builders.

By coupling open weights on Hugging Face with an enterprise customization backend in Tinker, Thinking Machines is positioning itself at the crossroads of community-driven open source development and enterprise cloud services. As developers begin benchmarking Inkling across production workloads, the debut establishes Murati’s venture as an active contender in the race to define the next generation of multimodal agent software.