FULL STORY

Perplexity's Open-Source Decider: From Release to Top of the Leaderboard

Perplexity open-sourced its multimodal decision model pplx-decider-v1-27b in early October, then shipped v1.1-27b, topping Hugging Face decision benchmarks at the lowest inference cost.

2026-10-02 ~ 2026-10-10 · 3 episodes · 27 posts

Episode 1 · Perplexity Open-Sources pplx-decider-v1-27b Decision Model (2026-10-02, 8 posts)

On October 2, Perplexity CEO Arav Srinivas announced the open-source multimodal decision model pplx-decider-v1-27b, with weights uploaded to Hugging Face (perplexity-ai/pplx-decider-v1-27b) and served through the new Decisions API at $0.04 per million input tokens with free output tokens. Unlike conventional language models, it produces probability distributions over fixed answer sets instead of text, positioning itself as a decision component for software systems.

Confirmed

  • The model is named pplx-decider-v1-27b, roughly 27B parameters. According to rm-rf-rm and inductionheads, it is fine-tuned from Qwen3 27B; jedisct1 relayed that it is based on Qwen3.8-27B with a 250k context window; inductionheads characterizes it as a "generative discriminator"
  • It supports multimodal understanding: per victormustar and Maziyar Panahi's inspection of the Hugging Face model card and chat template, the model handles image and video inputs
  • Per Perplexity researcher denisyarats, the Decisions API runs this model underneath, reading text, JSON, or images and returning structured answers with probabilities rather than generated text for direct software use
  • Pricing is $0.04 per million input tokens, with output tokens free
  • The announcement posts included no benchmark results or technical details

Why it matters

  • A decision model compresses LLM capability into programmable probabilistic output, removing free-text parsing and fitting directly into software workflows
  • The combination of open weights and a cheap API lowers the barrier for developers to try it in routing, classification, and selection tasks; victormustar sees it as another alternative in the emerging decision-model space (the Jev lineage)

Episode 2 · Perplexity Open-Sources Decider v1.1, Tops Both Decision Index and DecisionBench at Lowest Cost (2026-10-07, 16 posts)

Perplexity released the open-weight multimodal decision model pplx-decider-v1.1-27b on Oct 7, ranking first on the newly launched Hugging Face Decision Index 0.3; on Oct 9 CEO Arav Srinivas announced it also tops the open benchmark DecisionBench at the lowest cost. Weights are downloadable on Hugging Face, API input price is halved to $0.02 per million tokens (output free), and the model is available via OpenRouter.

Confirmed

  • Perplexity released pplx-decider-v1.1-27b, an open-weight multimodal decision model that scored highest on Hugging Face's Decision Index 0.3, with the API available simultaneously and listed on OpenRouter (relayed by AravSrinivas, perplexityai, gaganghotra and others)
  • Pricing halved versus v1: $0.02 per million input tokens, output free
  • In a clinical decision evaluation it answered 643 of 669 questions correctly, beating closed-source rival Jev's 628 at roughly 42% lower cost (announced by AravSrinivas, relayed by inductionheads and bowangbo)
  • On Oct 9, AravSrinivas announced Decider V1.1 ranked first on DecisionBench at the lowest cost: 93.9% accuracy and 534ms latency. DecisionBench is an open benchmark for small-model "bounded decisions" built from real records, covering engineering, agents, customer support, finance, legal and more, with reproducible accuracy/latency/cost metrics
  • Unlike typical generative models, it does not generate text: it reads text/JSON/image inputs and returns a typed decision with probabilities in a single request, suited for classification/routing tasks
  • Team member antoinechaffin said the model was iterated from v1 over a weekend sprint; the same backbone lifted the score from 56.4 to about 61.5 on Decision Index v0.2, about 3.5 points above then-leader Jev, before topping the v0.3 leaderboard
  • The model is publicly downloadable on Hugging Face and usable via a very low-cost API; antoinechaffin thanked multimodalart for building the evaluation and leaderboard and for running scores when the team submitted v0.2

Why it matters

  • Dedicated benchmarks for decision models are just emerging (Decision Index and DecisionBench), and Perplexity has simultaneously topped both with an open-weight model at the lowest cost, showing rapid iteration in this niche
  • Open weights, halved API pricing and multi-channel access via OpenRouter sharply lower the barrier to high-performance decision models, directly appealing to developers and cost-sensitive use cases

Episode 3 · Perplexity's 27B decision model tops DecisionBench at lowest cost (2026-10-09, 3 posts)

Perplexity released pplx-decider-v1.1-27b, a Qwen-based decision model that tops the DecisionBench leaderboard with 94.5% accuracy at the lowest cost, about $0.017 per thousand decisions.