✳ the wire · intel
Thinking Machines releases Inkling — open weights, 975B MoE, 1M context, multimodal
Thinking Machines' first from-scratch model is out with full weights on Hugging Face, plus an NVFP4 checkpoint: a 975B-total / 41B-active MoE, 1M-token context, natively multimodal across text, images and audio, pretrained on 45T tokens.
The positioning is deliberate and worth taking at face value. This is pitched as a broad, efficient base for fine-tuning — available on Tinker today at a limited-time 50% discount — rather than a flagship aimed at topping a leaderboard. Companies usually claim that when they have missed the frontier. Here the numbers support the framing rather than excusing it.
Every wire item is source-verified before it posts and graded: confirmed · strong · reported · rumor. Subscribers see the primary source on each one, the reasoning behind its grade, and the full analysis — plus the launch alerts themselves, which fire the minute a model answers rather than whenever the feed catches up.