Thinking Machines Lab released Inkling: a 975-billion-parameter model trained on 45 trillion tokens, fully open source, and benchmarked as a generalist across reasoning, coding, vision, audio, and factuality. No paint, no stereo. The weights are free.
The business model is the story. Inkling launched directly on Tinker, Thinking Machines' own fine-tuning platform. The base model costs nothing. The customization charges rent. It is the same logic Slate Auto used with its $24,950 electric pickup: hand-crank windows, no touchscreen, no speakers, unpainted composite body. Ship a stripped-down base, sell the accessories.
What makes this worth reading in full is the spider chart. Inkling's benchmark profile against Nemotron 3 Ultra, GLM 5.2, GPT 5.6 Sol, and Claude Fable 5 shows a model optimized for breadth, not dominance in any single category. That is a deliberate product decision, not a limitation. The question the original raises: whether this commercialization model, open weights plus proprietary tooling, becomes the standard playbook for US open-source AI.
[READ ORIGINAL →]