Overview
- Thinking Machines launched Inkling 2 Small as a proprietary model on its Tinker platform, making inference and on‑platform fine‑tuning available but not downloadable weights.
- Inkling 2 Small uses a mixture‑of‑experts design with 594 billion total parameters and about 17 billion active parameters per request, which cuts per‑call compute while preserving large capacity.
- The model is marketed for agentic coding, vision, reasoning, and knowledge work, accepts image inputs, and exposes a variable “thinking effort” control to trade latency and cost for deeper reasoning.
- Thinking Machines published per‑token pricing of $0.30 per 1 million input tokens and $1.00 per 1 million output tokens and released benchmark results including a 31 on the AAI index and fourth place on DesignArena.
- The company promises an eventual Apache 2.0 weights release after external safety testing but gave no date, a stance that will determine whether developers can self‑host, audit the model, or run it off‑platform.