Inklingが戦いの場に登場――「キャリブレーションされていない」モデルはむしろ強みか?
Mira Murati率いるThinking Machines Labは、推測する代わりに不確実性を示す975BパラメータのオープンウェイトモデルInklingを発表。独自データを取り込みながらカスタム価値を生まないプロプライエタリモデルへの支払いに疲れた企業がターゲットだ。
thinking machines lab launched inkling, a 975 billion parameter open-weight model that activates roughly 41 billion parameters per task to maintain efficiency. trained on 45 trillion tokens of text, image, audio, and video, the model focuses on calibrated responses and native reasoning across modalities rather than raw benchmark dominance. the company pitches this as a direct alternative to proprietary 'black box' models, allowing enterprises to fully customize and own their ai stack. this move aligns with predictions from industry leaders like microsoft's satya nadella and hugging face's clem delangue, who anticipate a shift from general frontier models to specialized, company-tuned open alternatives.