Perplexity open-sources Decider v1.3: 26B and 5B Apache-2.0 classification models with span extraction
Balance- · reddit · 2026-11-13
Perplexity released two open-weight (Apache 2.0) decision models: pplx-decider-v1.3 (fine-tuned from Qwen3.8-27B, 49 GiB BF16) and pplx-decider-flash-v1.3 (4B, KL-distilled from the 27B, 8.5 GiB). Instead of generating text, they take a state plus questions and return probability distributions over answer options.
Capabilities
- Yes/no, multiple choice (up to 255 options), ordered scores, and a new locate type returning verbatim spans with character offsets for NER, PII detection and extractive QA (up to 128 locate questions per request).
- Text and image input; multiple questions per state; 30k-token training sequences; bidirectional attention requires the custom pplxdecider package.
- Locate uses a separate unmerged LoRA adapter + span head, since merging regressed classification.
Reported results
- Multilingual classification (8 languages): 88.16% / 86.50% vs. 85.41% for TypeSafe Jev 1.14.
- Reranking over 72 retrieval tasks: 83.99 nDCG@10, +1.81 over Qwen3-Reranker-8B.
- Locate average over 17 out-of-domain tasks: 69.0 / 63.6 vs. 43.3 (Liquid d1) and 33.6 (GLiNER large v2.5).
- Artificial Analysis Decision Index 70 at $68/1M tasks and 64 at $34 — on the cost/score Pareto line.
Authors caution that probabilities should be calibrated and thresholded on your own data. Available on Hugging Face, OpenRouter and Perplexity's API.
Related event: Perplexity Open-Sources Decider v1.3 Decision Models(15 posts)→
More from Models
- Cloudflare hosts Thinking Machines' inkling 2 small on Workers AI at day-0 speed — michellechen · 2026-11-14
- Masked decoding trick tames text rivers in Gemma 3 12B, with a width-aware LoRA and coherence RL — cephaloform · 2026-11-14
- 250,000-human AI mafia game finds OpenAI models lie most, Anthropic snitches — Scobleizer · 2026-11-14
- Developer downgrades Claude Ultra to Pro and moves all work to Codex over 5-hour rate limits — umeshonai · 2026-11-14
- Pangram details per-domain false positive rates (1 in 10,000), enabling profitable AI detection — soldni · 2026-11-14
- Blogger predicts GPT-6.2 launches before Nov 19, floats three reasons for the delay — kimmonismus · 2026-11-14