Thinking Machines Unveils First Open-Weight Model, Inkling

Thinking Machines has unveiled Inkling, its first mission-related open-weight model. Designed with a Mixture-of-Experts architecture, the model has a total of 975B parameters (41B active) and supports a context window of up to 1 million tokens. It was pretrained on 45 trillion tokens including text, images, audio, and video.

Inkling-Small (12B active parameters) was also released, available for use cases requiring lighter and faster processing. Both models can be fine-tuned via the Tinker platform.

Key Features

• Multimodal reasoning: Natively processes text, images, and audio
• Efficient inference: "Controllable thinking effort" allows balancing cost and performance
• Broad capabilities: Competitive performance across diverse domains including agentic tasks, coding, reasoning, vision processing, and audio

Performance and Use Case Examples

In benchmarks, Inkling was designed as a balanced generalist model rather than optimized for a single domain. To achieve the same score, it requires only about 1/3 the tokens compared to Nemotron 3 Ultra. Real-world demonstrations included one-shot web app generation, multiplayer game development through 40 iterations of refinement, and multi-page PDF generation.

Tinker Platform Integration

Inkling can be tried directly by developers through the new Playground interface in the Tinker console. The article also showcased a self-retraining demo where Inkling itself used Tinker to write and execute fine-tuning jobs.