Thinking Machines Lab unveils Inkling 2 Small for enterprise workflows

Thinking Machines Lab, the artificial intelligence firm established by former OpenAI chief technology officer Mira Murati, has introduced its second-generation model, named Inkling 2 Small. The release targets organisations that want capable models for practical workplace tasks without being locked into closed, proprietary cloud systems.

The system uses a mixture-of-experts design with 594 billion total parameters, but activates only 17 billion parameters for any given token during operation. This setup is built to keep computational requirements manageable while processing complex inputs. The model handles text and images, supports a context window of up to one million tokens, and allows users to dial reasoning effort up or down depending on the task.

While currently available through the company's Tinker platform for testing and fine-tuning, the developer plans to release the model weights freely under an Apache 2.0 licence once external safety evaluations conclude. Third-party testing from Artificial Analysis gave the system an Intelligence Index score of 31, putting it ahead of recent open releases from European and Chinese competitors in the same tier.

The model was trained with a heavy emphasis on long-horizon reinforcement learning, combining coding, visual evaluation, and multi-step tool use into unified training runs. Benchmark evaluations released alongside the model show strong results on workplace simulations, including handling customer service inquiries, running internal software workflows, and writing code based on visual feedback.

What this means for Australian businesses

For Australian companies evaluating how to run AI tools inside their operations, the announcement brings several practical considerations:

• Open weights provide a clear path to data privacy. Once Apache 2.0 weights become available, businesses can run and fine-tune the system inside Australian cloud regions or on private servers, ensuring sensitive customer and financial records never leave local custody.
• Specialist tasks can be customised at lower ongoing fees. Because the model activates 17 billion parameters per token, hosting and inference demands are far lighter than running dense models of comparable scale, making internal custom tools more economical to run.
• Variable reasoning controls help manage daily running expenses. Staff can use fast, low-effort settings for standard data entry or email responses, then switch on deeper reasoning only when debugging code or assessing complicated spreadsheets.
• High-context visual skills open up practical internal automation. The ability to interpret images and long documents makes the tool suited to auditing invoices, checking digital design assets, and assisting internal support desks.