Thinking Machines Lab has released Inkling, a new multimodal mixture-of-experts model designed for token-efficient reasoning and native understanding of text, image, and audio inputs. Together AI is making the model available on its inference platform with day-zero access.
- Inkling features 975B total parameters with 40B active parameters per token and a 1M context window.
- The architecture uses query-conditioned relative attention, short causal convolutions, and a shared expert sink for efficient execution.
- It supports controllable inference effort, allowing developers to adjust reasoning depth, latency, and token usage.
- Preliminary evaluations show strong performance in scientific reasoning, mathematics, coding, and forecasting tasks.
- The model accepts text, image, and audio inputs through a unified decoder stack to produce text outputs.
Developers can access Inkling via Together AI Serverless with OpenAI-compatible APIs, enabling the creation of multimodal applications without managing infrastructure.