Thinking Machines Debuts Inkling Open-Weight AI Model
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Thinking Machines Lab released Inkling, an open-weight mixture-of-experts model with 975B total and 41B active parameters, trained on 45T tokens with a 1M context window
- Inkling handles text, image, and audio modalities using a sliding-window architecture with a 5:1 ratio and 512 size, and is available today for fine-tuning via Tinker
- VentureBeat frames the release around low cost and 'resistance to censorship,' an angle softened in most other outlets' coverage
- Financial Times reports that Inkling draws on techniques from Chinese rivals, notably DeepSeek-style aux-free load balancing with two shared experts
- Wall Street Journal positions the launch as Murati's bid to loosen the grip of established AI giants on the market
- Artificial Analysis scored Inkling at 41 on its Intelligence Index, calling it the new leading U.S. open-weights model
- Techstrong.ai casts the debut as a direct challenge to Chinese open-source dominance in frontier-tier open weights
Why it matters: Thinking Machines becomes the first U.S. lab outside the OpenAI-Anthropic-Google axis to ship a frontier-tier open-weight model, giving enterprise and academic users a fine-tunable alternative to closed APIs. The 41B active-parameter MoE design — borrowed in part from Chinese labs per the FT — signals that Murati's team is competing on the same architectural playbook as DeepSeek rather than trying to out-spend incumbents.
