FULL STORY

Perplexity Open-Sources Decider v1.3, Extends Free Access

Perplexity open-sourced its decision model Decider v1.3 with a one-week free run on OpenRouter, then extended free access by three more weeks as adoption grew.

2026-11-13 ~ 2026-11-20 · 2 episodes · 17 posts

Episode 1 · Perplexity Open-Sources Decider v1.3 Decision Models with One-Week OpenRouter Free Access (2026-11-13, 15 posts)

On November 13, Perplexity team member Antoine Chaffin announced open-weight decision models pplx-decider-v1.3 and the smaller, cheaper pplx-decider-flash-v1.3, amplified by CEO Arav Srinivas. Unlike generative chat models, Decider reads text, JSON, or image inputs and returns a typed, probability-scored structured answer (e.g. yes/no) in a single request without generating content. The models are live on OpenRouter, Hugging Face, and the official Decisions API.

Confirmed

  • v1.3 adds a locate capability that extracts specified spans from text by type or natural-language instruction, usable for PII detection. Per CEO Arav Srinivas, accuracy improved by 6.54 points across 63 production classification tasks, alongside performance and robustness gains; a detailed blog post was published.
  • Available on three channels: OpenRouter (fully free until November 19, handling both Decisions and Extractions requests), Hugging Face (Apache-licensed weights), and the official Decisions API.
  • Srinivas confirmed sizes: pplx-decider-v1.3 is 26B and Flash is 5B; weights are downloadable from Hugging Face for self-deployment, with Flash small enough to run locally.
  • Community user Balance reports pplx-decider-v1.3 is based on Qwen3.8-27B (49 GiB in BF16) and Flash is 4B (community-sourced, not official).
  • Perplexity engineer denisyarats showcased an internal security use case: Decider read 1 billion tokens across 7,000 agent sessions in one day, uncovering agent safety-rule flaws and yielding 14 fixes. Srinivas framed "Decision Models" as the future of agentic security, processing billions of tokens of telemetry at unprecedented scale and cost.

Why it matters

  • Decider represents a product philosophy distinct from chat models: no text generation, direct probabilistic answers suited to deterministic classification and extraction workloads.
  • The agent security audit case (1B tokens/day, 7,000 sessions, 14 fixes) is a concrete reference implementation and Perplexity's headline selling point.
  • Apache-licensed open weights (26B main + 5B Flash) plus a one-week OpenRouter free window lower adoption barriers; Flash's size enables local deployment.
  • Shipping with quantified production gains (63 tasks, +6.54 points) signals real-business iteration rather than a research prototype.

Episode 2 · Perplexity Extends Free Access to Decider Models on OpenRouter (2026-11-20, 2 posts)

Perplexity extended free access to its open-weight Decider-v1.3 and Decider Flash models on OpenRouter for another three weeks. CEO Arav Srinivas noted the model already accounts for nearly 20% of requests on the platform.