Perplexity Releases Decider v1.3 and Flash v1.3

Perplexity has released Decider v1.3 and Decider Flash v1.3, the latest versions of its decision models, in a blog post dated 12 November 2026. The company presents them as successors to Decider v1, v1.1, v1.2 and Flash v1.2. TSN previously covered the v1.1 release alongside OpenAI’s Decisions API.

The blog says Decider Flash v1.3 is “a smaller model distilled from Decider v1.3.” It also says: “Both models are available on Hugging Face under Apache license, through OpenRouter, and the Perplexity Decisions API.” That Hugging Face and Apache availability is Perplexity’s account in the blog. At the time TSN checked the model cards on 13 November 2026, the cards for pplx-decider-v1.3 and pplx-decider-flash-v1.3 each said: “The repository is private: authenticate with a Hugging Face token that has access to it.” This is not a fully open download.

What the new models do

The v1.3 model card describes pplx-decider-v1.3 as a multimodal classification model built on Qwen3.8-27B, and the next version of pplx-decider-v1.2. It says a CUDA GPU needs room for about 49 GiB of BF16 weights, plus working memory. The Flash card describes pplx-decider-flash-v1.3 as the small version of the same family, built on Qwen3.5-4B, distilled from the 27B model, and needing about 8.5 GiB of BF16 weights plus working memory. Both cards list an Apache 2.0 licence in their metadata.

Perplexity says Decider supports binary and multi-class classification, with ordered or unordered labels, and returns probabilities that applications can use to classify, route, rank and evaluate inputs. The v1.3 card says a question can have up to 255 options. The blog says both v1.3 models improve performance and robustness across a range of text and vision tasks “through bidirectional attention, better-curated training data, and a refined training recipe.”

The other addition is a feature Perplexity calls locate. The blog says it is “a new feature to highlight the selected content for named entity recognition, personal information identification, and extractive question-answering tasks.” The v1.3 model card adds that Locate extracts verbatim text spans from the state, with character offsets and confidence scores.

The blog says Decider is inspired by Jev, TypeSafe AI’s System One Model, a line TSN has also followed, including when Vercel reported Jev as the fastest-adopted model on its AI Gateway. A separate, more recent decisions interface from another vendor is Google’s Gemini Decisions API, launched with Gemini Decision 1.

Perplexity’s getting-started advice is that Decider returns probabilities, and that the right decision threshold depends on the application, so users should tune thresholds on a small, labelled development set rather than assume a fixed cut-off. It recommends Decider v1.3 for nuanced judgements, detailed rubrics and workflows where errors are costly, and Decider Flash v1.3 for high-volume routing, labelling and filtering where latency and cost matter more. It tells teams to evaluate both models on their own workload. The cards also say training used sequences of up to 30,720 tokens, and that inputs are not truncated, so longer inputs should be validated on the user’s own task.

Analysis

The practical question for anyone who wants to run these models is not only what they do, but how they can be obtained. Perplexity’s blog states that both models are available on Hugging Face under an Apache licence, as well as through OpenRouter and the Perplexity Decisions API. The model cards TSN opened on 13 November 2026 said the repositories were private and required a Hugging Face token with access. Those two statements can both be true at once: a company can announce Apache-licensed weights on Hugging Face and still gate the repository behind an access request. Until that gate is lifted, or Perplexity says otherwise, it is safer to treat the blog as the claim of availability and the card as the live access condition, not to describe v1.3 as a fully open download.

The versioning is clearer than the access wording. This is an incremental step from the v1.1 models TSN recorded earlier, through v1.2, to v1.3, with Flash again distilled from the larger model. The cards’ backbones — Qwen3.8-27B for Decider v1.3 and Qwen3.5-4B for Flash — and the roughly 49 GiB and 8.5 GiB BF16 weight sizes are the hardware facts to plan around, not marketing labels. A classification head that returns probabilities, rather than a free-form paragraph, is the same product idea as the earlier Decider and as the newer Gemini Decisions API: applications can set a threshold, route a ticket, or decline to act when the probability is weak.

Locate is the part that most obviously extends that idea. Extracting a verbatim span, with character offsets and a score, is closer to named-entity recognition or extractive question answering than to a yes-or-no decision. It still depends on the same caution as the probability scores: Perplexity is describing a capability and a recommended way to tune thresholds, not publishing a TSN-verified accuracy figure in this article. Teams that handle personal information, in particular, will want to test that feature on their own examples before they treat a highlighted span as a redaction or a fact.

As with other decision-model launches, the useful test is a representative labelled set from the actual workload, not the existence of a blog, an API, or a model card. Flash exists for people who need a smaller, distilled model; v1.3 exists for people who want the larger one. Which of those is “better” is a question the blog itself declines to answer in the abstract, and this article should do the same.