TypeSafe AI has announced the availability of Jev 1.14, the latest iteration of its System One decision model, now accessible under the jev-preview alias. This update primarily introduces three experimental features: the ability to process images, a dedicated 'Box' output type for spatial understanding, and a substantially increased context window. These additions are designed to propel the model towards sophisticated vision-based tasks, such as object inspection and localization, for developers.
Enhanced Context and Core Model Performance
Jev 1.14 maintains the default context window at 64,000 tokens per request and 32,000 tokens per question. However, developers can opt-in to an extended context of 262,000 tokens by including a specific header. The model operates exclusively via TypeSafe's hosted API, without publicly disclosed parameters or open-source weights. Internal assessments suggest a general improvement across the model, with the most notable gains observed in multilingual processing capabilities. TypeSafe's 'Box test,' an internal evaluation, demonstrated the model's ability to identify 30 out of 36 partially cropped targets using four edge-checking questions, showcasing its potential for precise localization.
Jev, at its core, is distinguished by its ability to return typed decisions rather than generative text. Users submit a state alongside structured questions (Choice, Score, or Noul), and Jev responds with typed values, probabilities, and confidence scores. The model is trained using Reinforcement Learning for Calibrated Decisions (RLCD).
Integrating Visual Input with Images
The new image input functionality, while experimental, marks a significant step. Images can be incorporated by referencing them within the model's state, instructions, or criteria, allowing for flexible integration. This feature requires activating a specific 'danger zone' header in each request. Jev supports static PNG or JPEG images, with individual files limited to 16 MiB and dimensions between 256 and 1568 pixels per side. The total image data per request cannot exceed 32 MiB. Images are billed as standard input tokens, without additional fees. TypeSafe cautions that the model's performance with image inputs currently lags behind its text-only capabilities, advising users to formulate the simplest possible questions when working with visual data.
Introducing the 'Box' Question Type
A novel 'Box' question type has been implemented, designed to output a single bounding box per query. This answer provides normalized top, bottom, left, and right edge coordinates (ranging from 0 to 1), accompanied by a confidence score. The 'Box' type necessitates Jev 1.14 (or jev-preview) and the image input header. A crucial aspect to note is that the 'Box' type consistently returns coordinates, even if the target object is absent. TypeSafe recommends coupling 'Box' questions with a 'Noul' type question to confirm the existence of the desired object.
Availability, Pricing, and Future Outlook
The Jev 1.14 update is currently available through the jev-preview alias, with a transition to jev-latest expected within a week. The previous version, Jev 1.13, will remain accessible with a 60-day deprecation notice. Pricing remains consistent at $0.042 per million input tokens, with output tokens being free of charge. While TypeSafe has published internal performance charts comparing Jev against open-source models, these do not include specific numerical scores and are acknowledged to have a potential bias towards their proprietary models. Developers are encouraged to conduct their own evaluations.
These new features, particularly image processing and the 'Box' type, are explicitly flagged as experimental and unsuitable for production deployments at this stage. Users are advised to exercise caution and prioritize shorter contexts for optimal reliability.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
