Product Launch3 min read

Perplexity releases Decider v1.3 classification models with open weights

Perplexity has released Decider v1.3 and a smaller Flash version under the Apache license on Hugging Face. The models handle classification, scoring, reranking and now span extraction, and can run on a team's own hardware.

AE
Aidose Editorial
Nov 16, 2026

Perplexity has released Decider v1.3 and Decider Flash v1.3, the latest versions of its open-weight decision models, according to a November 12 post from Perplexity Research. Both models are available as downloadable weights on Hugging Face under the Apache 2.0 license, and through OpenRouter and the Perplexity Decisions API.

Decider is built for structured decisions rather than chat. Per the Hugging Face model card, each request uses one of four types: choice, which picks among 1 to 255 named options; a yes/no type that returns the probability a statement is true; score, which places input on a scale of 2 to 255 levels; and locate. Locate is new in v1.3 and extracts exact text spans with character offsets, for tasks such as entity recognition, PII detection and extractive question answering.

The larger Decider v1.3 is fine-tuned from Qwen3.8-27B and needs about 49 GiB of BF16 weights. Flash v1.3 is distilled from it into a Qwen3.5-4B base. Perplexity says its full-attention layers use bidirectional attention, and that each question attends only to the shared input and its own tokens, so many questions can be asked about one document at once. Training used sequences up to 30,720 tokens.

Perplexity reports 91.6% mean accuracy across 63 query-classification tasks for v1.3, which it says is 6.54 points above its own production classifiers. It reports 83.99 mean nDCG@10 across 72 retrieval tasks, 1.81 points above Qwen3-Reranker-8B. On span extraction, v1.3 averaged 69.0 and Flash 63.6, against 33.6 to 43.3 for the comparison models. The answer flip rate when labels were reworded was 0.27%. These are Perplexity's own benchmarks.

For teams that route queries, moderate content, rerank search results or redact PII, Decider gives them a single permissively licensed model to run on their own servers instead of calling a general-purpose LLM for each check. The 4B Flash variant is small enough to deploy cheaply, and Perplexity recommends the 27B model for nuanced rubrics and workflows where errors are costly.

AE
Aidose Editorial
Nov 16, 2026 · 3 min read
Back to News