Thinking Machines Lab has released Inkling, its first model trained from scratch with open weights available for fine-tuning. The model is a Mixture-of-Experts transformer featuring 975B total parameters and 41B active parameters, supporting a context window of up to 1M tokens.
- Pretraining utilized 45 trillion tokens of text, images, audio, and video data.
- It accepts text, image, and audio inputs but outputs UTF-8 text only.
- The architecture includes a 66-layer decoder-only transformer with sparse MoE feed-forward layers following the DeepSeek-V3 design.
- Inkling features controllable thinking effort, allowing users to adjust token budgets via system messages or API arguments.
- It spends one-third as many tokens as Nemotron 3 Ultra for equal Terminal Bench 2.1 performance.
- A smaller variant, Inkling-Small (276B total, 12B active), is in testing and will release its weights later.
The model offers tunable cost and latency per call through its controllable effort mechanism, positioning it as a base for customization and agentic pipelines.