Perplexity has open-sourced two decision models. The 26B version requires about 49 GiB of memory, while the 5B lightweight version requires roughly 8.5 GiB; both use the Apache 2.0 license. Our judgment is that models that do not generate free-form text and instead return direct judgment probabilities are becoming a layer worth provisioning separately within enterprise AI infrastructure.
What this is
Decider v1.3 is a classification and decision model based on the Qwen backbone model. It takes a state description and a set of questions, then returns yes/no answers, multiple-choice selections, ranking scores, or the positions of text spans instead of generating an answer word by word like a chat model. Its new locate function can extract spans such as names and email addresses from source text and provide their starting and ending character positions.
The 26B version offers higher accuracy, while the 5B Flash edition reduces resource requirements through distillation—training a smaller model on the answer distributions of a teacher model. Both support text and images, and both are available through APIs via Perplexity and OpenRouter.
Industry view
What is encouraging is that, across 17 cross-domain tests published by Perplexity, the large model achieved an average score of 69.0 and Flash scored 63.6, outperforming the Liquid d1, Cloudflare Clairic, and GLiNER models listed by Perplexity. This matters because classification, reranking, and sensitive-information detection do not all need to rely on costly large-model inference; enterprises can combine models in layers according to accuracy and cost.
The risks are equally clear: these results come primarily from Perplexity’s own testing and do not yet demonstrate equivalent stability in healthcare, finance, or complex document processing. In batch queries, different combinations of questions may slightly alter the answers. Requests that mix locate and classification operations also require additional computation. Open weights do not mean that the models are ready to use out of the box: their bidirectional attention architecture still depends on a specialized runtime, and enterprises must bear the costs of evaluation and operations.
Impact on regular people
For enterprise IT: Classification models may become low-cost components in systems for content moderation, ticket routing, and contract-field extraction. They are especially suitable for organizations that require data to remain within their own environments, but real business samples should still be used for retesting before deployment.
For individual work: AI tools will become better at prioritizing work, checking materials, and identifying sensitive information, rather than merely drafting text. Workflow designers will need to define rules and acceptance criteria more clearly.
For the consumer market: Locally runnable models will push photo classification, receipt organization, and privacy screening onto personal devices. However, differences in speed, memory consumption, and device capabilities will still limit large-scale adoption.