Overview
- Inkling, released July 15, is an open‑weights mixture‑of‑experts model with 975 billion total parameters that activates about 41 billion per task and natively handles text, images, audio and video.
- Thinking Machines trained Inkling from scratch on a reported 45 trillion tokens, then used larger‑scale reinforcement learning for post‑training after initially bootstrapping some early post‑training data with other open‑weight models.
- The full weights are available for download on hubs such as Hugging Face and the model launched with day‑zero commercial partners including Databricks through its Unity AI Gateway and several inference and hosting providers.
- Company benchmarks show Inkling is competitive on agentic and coding tasks but the lab acknowledges Inkling is not the top model overall, and it offers controllable “thinking effort” and long‑context modes to trade cost for performance.
- The release shifts commercial dynamics by letting firms avoid per‑token APIs and keep proprietary data in‑house, which could boost enterprise customization but forces Thinking Machines to rely on services, tooling and hosting for revenue and may reshape how U.S. companies compete with stronger foreign open models.