Perplexity has released 'Decider v1.3' and 'Decider Flash v1.3', open-weight multimodal decision models that perform classification, routing, scoring, and information extraction in a single pass.
Imagine this: In a large-scale customer service center flooded with thousands of customer inquiries, documents, and images, staff members are exhausted from constantly classifying and scoring whether each request is a refund request, a simple inquiry, or an urgent complaint. Now, an AI has emerged that handles this complex and tedious judgment task in just a few seconds, with greater accuracy than humans. This is the story of “Decider v1.3” and “Decider Flash v1.3,” open-weight (public weight models whose internal structure anyone can examine and utilize) multimodal (technology that simultaneously understands various forms of data such as text, images, and audio) decision models released by Perplexity (an AI-based search and conversational platform) [Source 2, Source 7].
In this article, let’s take a step-by-step look at why this new decision-making AI from Perplexity is gaining attention and what changes it will bring to our daily lives and businesses.
Why is this important?
The most troublesome point for businesses when adopting AI is “judgment and decision-making.” By analogy, it is like a massive factory standing still because workers are manually inspecting countless parts one by one for defects. To answer questions like “Is this email spam?”, “Did the documents submitted by this customer meet regulations?”, or “Is this product review positive or negative?”, companies previously had to complexly wire together different programs.
However, Perplexity’s Decider v1.3 achieved an average accuracy of 91.6% across 63 query-classification (the task of categorizing user questions or request intents by type) tasks, proving a performance 6.54 percentage points higher than its existing proprietary production classifier [Source 2].
Going beyond simply reading and answering text, it can process classification, scoring, and extraction in a single pass [Source 4, Source 6], enabling companies to build fast and accurate automation systems at a much lower cost.
Easy to Understand: How Does Decider v1.3 Work?
Simply put, Decider v1.3 can be compared to a “high-speed judge” or a “genius librarian in a large library.”
This model simultaneously ingests various forms of data, including text, structured JSON (a lightweight text format for storing and exchanging data), and even images [Source 6]. It thoroughly evaluates them based on yes/no questions, custom user-defined choices, or ordered rubrics (evaluation criteria tables), and then outputs typed answers along with probability values [Source 6].
Core Features Through Analogities
- X-ray Security Scanner: Just as an airport scanner inspects luggage, Decider v1.3 detects regulatory violations or specific items within vast amounts of data without error.
- New ‘Locate’ Feature: The ability to pinpoint specific information directly within text has been added [Source 1, Source 7]. It is a technology that finds specific words as precisely as cutting them out with a knife from a massive encyclopedia.
The biggest difference from existing AI is that all these processes occur organically within a single model [Source 4, Source 6].
Current Status: Where Do We Stand?
Perplexity Decider v1.3 is the latest model born from the stellar lineage of previous versions v1, v1.1, v1.2, and Flash v1.2 [Source 2]. The technical features and specifications identified so far are as follows:
- Powerful Multimodal Processing: It can simultaneously evaluate not only text but also images and structured data [Source 6].
- Massive Context Window: It provides an enormous context window (the amount of information AI can remember and process at once) of 256K (approx. 256,000 tokens), allowing it to ingest and analyze an entire book’s worth of data at once [Source 6].
- Lossless Preservation of Input Data: During training, it used sequences up to 30,720 tokens (the basic unit by which AI reads and processes sentences) in length, adhering to the principle of never truncating input data midway [Source 9].
- Flexible Question Limits: There is no fixed limit on the number of questions, allowing numerous questions to be posed at once within the allowed length limit [Source 9].
Developers and enterprises can easily test this model and use it via APIs through OpenRouter or Vercel AI Gateway [Source 6, Source 8].
What’s Next?
As Perplexity fully open-sources Decider v1.3 and Decider Flash v1.3 as open-weight models [Source 2, Source 7], countless companies and developers are expected to implant this “intelligent judgment engine” into their respective services.
In areas where “accurate judgment” determines business success or failure—such as customer center automated classification systems, large-scale contract and legal document review, and real-time content moderation—Decider v1.3 is poised to become core infrastructure. The clock hand is moving fast as AI evolves beyond simply speaking well into becoming a true “expert advisor who gets things done.”
AI’s Perspective
MindTickleBytes AI Reporter’s Perspective: AI’s advancement is evolving beyond simply writing well into the precision of making “correct judgments.” Decider v1.3 proves that AI is not just a chatterbox, but can become a reliable advisor working alongside business’s most demanding decision-makers.
References
- Perplexity Decider v1.3: Training, eval, and use cases
- Perplexity Ships Decider v1.3 Classifier, 91.6% on 63 Tasks
-
[Perplexity’s Decider v1.3 lets one model · 21 HN comments Zeli](https://zeli.app/story/50488464) - Decider v1.3: Release, Decision Index Score, Performance …
-
[Decider V1.3 API, Pricing & Playground Vercel AI Gateway](https://vercel.com/ai-gateway/models/pplx-decider-v1-3) - Perplexity Open-Sources Decider v1.3 and Decider Flash v1.3 …
-
[DeciderV1.3(free) - API Pricing & Providers OpenRouter](https://openrouter.ai/perplexity/pplx-decider-v1.3:free) - perplexity-ai/pplx-decider-v1.3 · Hugging Face
-
[Perplexity Decider v1.2 27B Benchmarks & Pricing BenchLM.ai](https://benchlm.ai/models/pplx-decider-v1.2) - How To Use Decision Models for Evals – Hamel’s Blog
-
[Perplexity Decider V1.2 API Pricing & Benchmarks LMSpeed](https://lmspeed.net/model/pplx-decider-v1.2)
- It can only process text
- It performs classification, scoring, and information extraction in a single pass
- It always truncates input data midway
- 81.2%
- 91.6%
- 99.4%
- Locate
- Translate
- Compile