Inkling: Thinking Machines’ Open-Weights Model

What it is

• Inkling is a mixture-of-experts model from Thinking Machines, the company founded by former OpenAI CTO Mira Murati.
• It has 970 billion total parameters, but activates only about 41 billion per token by routing each token to a subset of specialized experts.
• It was trained from scratch on 45 trillion text, image, and audio tokens and supports a 1-million-token context window.
• Its weights are released under the Apache license and available on Hugging Face.

What makes it notable

• Adjustable thinking effort: Users can trade speed and cost for deeper reasoning. The video says Inkling matches Nemotron 3 Ultra on Terminal Bench at high effort while using roughly a third of the tokens.
• Fine-tuning through Tinker: Thinking Machines’ Tinker API lets users fine-tune open models without managing training infrastructure.
• Self-modification demo: Inkling used Tinker to write a training script, create data, train a new version of itself, and load weights that prevented it from using the letter “E.”
• Uncertainty awareness: Training rewarded the model for admitting when it did not know an answer. The video says this helped it perform strongly at forecasting future events.
• Direct multimodal input: Rather than first converting images and audio through separate encoder models, Inkling processes raw pixels and audio directly.

Limitations and trade-offs

• Inkling is not presented as the world’s strongest model. The video says it trails leading frontier models on benchmarks and faces strong competition from other open models.
• Extended reinforcement learning led its internal reasoning style to become terse, sometimes dropping words to save tokens.
• Its intended advantage is not raw intelligence, but efficient inference and specialization: start with an open model, then fine-tune it for a particular task.

Business strategy

Thinking Machines appears to be positioning Inkling as a capable, freely available base model and Tinker as the tool for turning it into a specialist—rather than trying to beat frontier labs across every benchmark.

Sponsor: Clerk

The video also promotes Clerk, which offers a CLI, agent skills, and an MCP server for setting up authentication, billing, and user management through coding agents. The presenter says an agent created auth pages and middleware, added paid plans, and retrieved user data without manual dashboard work.