AI ·
Google Challenges Jev With Multimodal Decision Model
Google has introduced Gemini Decision 1, a model designed to evaluate choices for AI agents rather than generate text. It leads Jev on five of seven shared benchmarks, while Jev retains advantages in tool calling and multimodal web tasks.
Google has entered the decision-model market with Gemini Decision 1, positioning it as a challenger to TypeSafe Jev two months after that model helped popularize the category. The launch also brings fresh attention to Jev’s $7.5 billion valuation, which had previously prompted questions about whether investor enthusiasm was justified.
Like Jev, Gemini Decision 1 produces probabilities and confidence scores for candidate actions instead of written answers. Software can use those outputs to guide an agent directly, avoiding token-by-token answer generation and the associated output-token charges.
The model can assess as many as 1,000 independent decisions within one shared context. It accepts mixed inputs spanning text, images, audio, video and PDFs, allowing agents to combine judgments that might otherwise require separate model calls. Google’s approach targets recurring agent work such as reviewing documents, selecting tools and checking task progress.
Across seven benchmarks reported by both sides, Gemini Decision 1 posted better results on five. Jev still led on When2Call for tool calling, scoring 82.0% against Google’s 79.9%, and narrowly retained its Mind2Web multimodal advantage at 86.8% versus 86.2%.
Source: 雷峰网
Written with AI from the linked source