Perplexity’s New Decider Models Turn Text Into Clear Decisions

Perplexity released Decider v1.3 and Decider Flash v1.3 on November 12, 2026, with a practical upgrade: the models can now pull out exact text spans instead of only returning decisions. The company also made downloadable Apache-2.0 weights available, giving teams the option to use the models through a service or run them themselves.

The release targets applications that need reliable answers to focused questions, not long pieces of generated prose. Perplexity reports 91.6% mean accuracy for the larger model across 63 query-classification tasks, with references independently relabeled by a frontier LLM. That score sits 6.54 percentage points above Perplexity’s production classifiers.

Locate adds exact text extraction

The new Locate feature returns matching text with its start and end character positions. It supports named-entity recognition, personal-information detection and extractive question answering, so an application can identify the precise part of a document that answers a question or triggers a rule.

When there is no match, Locate returns an empty list. The feature remains text-only and allows up to 128 Locate questions per request. Locate is available through both models, OpenRouter and Perplexity’s Decisions API.

Decider’s other output is numerical. It returns probabilities for yes-or-no questions, custom choices or ordered scores instead of generating prose. An application can ask several questions about the same text, JSON or images, then use the numbers for classification, request routing, moderation, filtering, output evaluation or reranking. These are jobs that often need a decision threshold rather than free-form generation.

How Decider v1.3 changes the model design

Decider adapts generative backbones by replacing next-token output with a decision head read at a marker token. That head starts with only the base-model rows associated with answer-option tokens.

For v1.3, Perplexity makes the full-attention layers bidirectional and stops questions from directly attending to one another. The shared input state can still attend to the questions. This replaces v1.2’s arrangement, where an answer could change depending on what else was asked, and the new design is intended to make answers more stable.

Training converts hundreds of public datasets into a shared state, question, options and target format. Generalist fine-tuning excludes benchmark training splits and removes samples that Perplexity says could overlap its evaluations. Later domain and vision specialization uses LoRA, which freezes pretrained weights while training injected rank-decomposition matrices.

Perplexity uses weighted adapter merges to control drift and averages weights from three runs for the text stages. The company then distills Flash from the larger model, using KL divergence to reduce the difference between its output probabilities and the teacher’s rather than learning from gold labels alone.

Model size, hosting and request limits

The Decider v1.3 model card lists a Qwen3.8-27B backbone and roughly 49 GiB of BF16 weights. Decider Flash v1.3 uses Qwen3.5-4B and requires roughly 8.5 GiB for its weights. Self-hosting either model requires a CUDA GPU and working memory beyond the weights, while Flash offers the smaller hardware footprint.

Locate has an important operating tradeoff. Requests that mix Locate with decision questions require separate forward passes for each type, adding latency compared with requests containing only one type. Perplexity keeps the Locate adapters unmerged because merging them caused regressions on ordinary decision requests, and the company says those adapters activate only for Locate requests. That preserves existing decision behavior, but mixed workloads cannot fully share the same computation.

A growing field of decision models

Perplexity’s release arrives alongside new decision-focused tools from Google DeepMind, OpenAI and Amazon. Google DeepMind released the Gemini Decisions API Wednesday, with Gemini Decision 1 able to answer up to 1,000 questions about the same material in one pass. Google charges $0.10 per million input tokens, with free output, and lists the model through the Gemini Developer API’s v1beta endpoint, Google AI Studio, OpenRouter and Vercel AI Gateway.

Google’s launch benchmarks show Gemini Decision 1 leading five of the seven tests that all four models ran. It scored 92.2% on JevBench, compared with 90.9% for GPT-6 Luna and 90.0% for Jev 1.14. Jev led on Mind2Web Multimodal, 86.8% to 86.2%, and on When2Call, 82.0% to 79.9%.

On the media tests, Decision 1 scored 71.8% to 67.7% against Gemini 3.5 Flash-Lite on MMAU, and 69.1% to 63.3% on Video-MME. It also had the lowest calibration error in Google’s comparison at 0.0250, versus 0.0494 for Jev. Google reports shared-test costs of $0.88 with Decision 1, $0.68 with Jev and $4.14 with Gemini 3.5 Flash-Lite through the Interactions API.

OpenAI made its Decisions API generally available on Oct. 22. GPT-6 Luna is the sole supported model, with requests supporting up to 400 questions and evidence from text, images, native JSON or previous tool calls. OpenAI charges $0.10 per million uncached input tokens and $0 for output, while cache reads cost $0.01 per million tokens.

Amazon released Strands Decider 2B in October. Its Qwen3.5-based 2B checkpoint supports yes-or-no, multiple-choice and rubric questions over text, runs on a CPU or GPU, and is available under Apache 2.0 with serving code and a training recipe.

For developers, the key difference is not only model accuracy. Each system also brings its own limits, hardware needs, pricing and integration work. Perplexity’s Decider v1.3 adds exact text extraction to a decision system, while Flash offers a smaller self-hosting option. That combination gives teams another way to turn documents, images and structured data into decisions they can use directly.

Stay connected via Google News

Follow us for the latest AI News