Thinking Machines Launches 975B Open-Weight Inkling
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Thinking Machines Lab debuted Inkling, an open-weight mixture-of-experts model with 975B total and 41B active parameters, trained broadly rather than optimized for a single area
- Inkling spans text, image, and audio modalities with a 1M-token context window, 45T training tokens, and DeepSeek-style aux-free load balancing using 2 shared experts
- Mira Murati released Inkling's full open weights on day one alongside a Hugging Face model card and an Inkling Playground, positioning the model as a counter to 'one-size-fits-all' AI
- John Schulman said pretraining on Inkling began last winter and a small team scaled up coding, reasoning, and agentic training from mid-January
- Outlets including the Wall Street Journal, Financial Times, and Techstrong.ai frame Inkling as Thinking Machines' bid to loosen closed AI labs' grip and challenge Chinese open-source dominance
Why it matters: Releasing a 975B-parameter open-weight model on day one with full fine-tuning access via Tinker gives developers a frontier-scale model to build on without API lock-in. WSJ, FT, and Techstrong.ai frame it as a dual challenge to U.S. closed labs and Chinese open-source rivals, though Thinking Machines' own launch stresses broad capability over ideological positioning.
