Thinking Machines releases Inkling 2 Small model

Thinking Machines has unveiled Inkling 2 Small, its second generation of models, according to a report by The-decoder. The startup, led by former OpenAI CTO Mira Murati, describes the mixture-of-experts model as having 17 billion active parameters. The new model aims to serve as a base for fine-tuning specialised applications. Thinking Machines intends to make the weights available under an Apache 2.0 licence following external safety testing.

Why it matters

The release of Inkling 2 Small suggests a shift toward more efficient, specialised automation. For businesses, the move toward Apache 2.0 licensing means you may have more freedom to integrate these models into proprietary workflows without the restrictive terms found in Meta or Mistral models. However, the increased token usage per task requires a careful cost-benefit analysis.

• Integration costs will depend on the complexity of your multi-step agentic workflows.
• Fine-tuning the model allows for higher accuracy in niche sectors like legal or technical documentation.
• The trade-off between speed and thoroughness gives you more control over operational latency.

What to watch

You should monitor the model's performance in multi-step workflows before committing to large-scale deployments. While the cost per task remains lower than Gemini or Claude, the higher token output per task can impact your budget unpredictably. If you plan to use the model for coding or visual reasoning, test its ability to select the correct tools for your specific codebase.

Within a year, Inkling 2 Small will likely become a standard for local deployment. This requires hardware providers to reduce the latency penalty of running these models on consumer-grade chips.