level: business
thinking machines lab, founded by former openai cto mira murati, released its first proprietary ai model called inkling. it is an open-weight mixture-of-experts system with 975 billion total parameters, activating about 41 billion per task. trained on 45 trillion tokens of text, image, audio, and video, it reasons across modalities. the company positions it as a starting point for organizations to fine-tune using its tinker platform, rather than a finished product.
the model emphasizes calibrated answers, flagging uncertainty instead of guessing, and lets users adjust thinking effort. on coding benchmarks, it uses a third of the tokens of nvidia's nemotron 3 ultra for similar performance. thinking machines does not claim inkling is best-in-class, but aims for well-rounded performance. the release follows a research preview of interaction models that listen and speak more naturally than typical chatbots.
the company argues that centralized, one-size-fits-all models underperform compared to ai that organizations customize themselves. this view is echoed by microsoft ceo satya nadella and hugging face ceo clem delangue, who see a shift toward private or open-source models for production. a project with bridgewater associates showed a customized open-source model beating proprietary ones on financial reasoning at lower cost. thinking machines reached this stage in about nine months, faster than competitors, though it partly used other open-weight models for early post-training data.
why it matters: open-weight models like inkling let enterprises customize ai without vendor lock-in, potentially reducing costs and improving performance on specialized tasks.