TypeSafe AI ships Jev 1.14.0 with image input and 262k-token contexts

Jev has drawn attention as a 'decision model' designed on very different lines from large language models, and Jev 1.14.0 is its latest iteration. According to TypeSafe AI, Jev-1.13.0 was multimodal from the start, but the company disabled that capability because models tend to be less intelligent on modalities other than text. Image support in Jev 1.14.0 is configured through a retrieval-replacement function, and the company cautions that questions are best broken down into their simplest elements.

The release also reflects a longstanding tension TypeSafe AI says it has discussed for years: balancing reliability and creativity. The company argues that models shipped to product builders should meet internal quality standards and be stress-tested to the limits possible in-house, but that imposing an extreme reliability standard would not produce great products. It is therefore sharing new features under a 'danger zone' API header flag.

Alongside the model changes, TypeSafe AI added an opt-in feature letting the community help find inconsistencies using queries, which it says makes problems easier to spot and use cases easier to debug. Its developers remain skeptical of public benchmarks, and the company's own comparison of Jev 1.14.0 with previous models and other decision models comes with the CEO's warning that the numbers are likely biased toward its own models.