Descriptions:
Fireship covers the debut of Inkling, the first major model release from Thinking Machines — the AI lab founded by Mira Murati after her departure as CTO of OpenAI. The video breaks down what makes Inkling technically distinct and explains the business logic behind Thinking Machines’ open-weights strategy.
Inkling is a mixture-of-experts model with 970 billion total parameters, but only 41 billion activate per token, delivering frontier-scale reasoning at reduced compute cost. It was pre-trained on 45 trillion tokens of text, images, and audio, handles a one-million-token context window, and is released under an Apache 2.0 license on Hugging Face. Unlike most multimodal models, Inkling processes raw audio and pixels directly rather than routing them through separate encoder models. A “thinking effort” dial lets developers trade speed and cost against reasoning depth — at high settings, the model reportedly matches Neotron 3 Ultra on Terminal Bench while using a third of the tokens.
Fireship also covers Inkling’s “epistemics” training — a reinforcement learning approach that rewards the model for expressing uncertainty rather than confabulating — which gives it strong performance on future-event forecasting, reportedly beating GPT 5.5 and Opus 4.8. The business model is straightforward: give away a capable open-weights base model and charge for fine-tuning through Tinker, Thinking Machines’ managed training platform. The video notes that Inkling’s launch was immediately overshadowed by Moonshot’s Kimmy K3 announcement, framing the competitive timing as a challenge for the new lab.
📺 Source: Fireship · Published July 20, 2026
🏷️ Format: News Analysis
