Tools & Products

Perplexity Open-Sources Decider v1.2 and Decider Flash v1.2: 27B and 4B Multimodal Decision Models

Perplexity AI open-sourced weights for Decider v1.2 (27B) and Decider Flash v1.2 (4B). The models specialize in sub-20ms query routing, confidence calibration, and real-time retrieval dispatch.

By Julian Thorne · 2026-10-18 · 8 min read
Perplexity Decider Flash v1.2 social card highlighting open-source weights and multimodal decision routing
Perplexity Decider Flash v1.2 social card highlighting open-source weights and multimodal decision routing

On October 18, 2026, Perplexity AI open-sourced weights for two dedicated decision models: Decider v1.2 (27 billion parameters) and Decider Flash v1.2 (4 billion parameters).

Unlike general conversational models that produce free-form text paragraphs, decision models output structured classification vectors, routing decisions, and calibrated probability distributions. Perplexity uses these systems internally to determine when to trigger web crawlers, invoke calculator tools, or route traffic to frontier reasoning clusters.

Technical Specifications and Footprint

Both models are published under the Apache 2.0 license, permitting commercial integration into proprietary software systems:

Technical Spec Decider Flash v1.2 Decider v1.2 Generic 70B LLM (Baseline)
Active Parameters 4.1 Billion 27.2 Billion 70 Billion
Target Function Real-Time Routing High-Dimensional Intent Class General Chat
Inference Latency (L4 GPU) 14 Milliseconds 48 Milliseconds 320 Milliseconds
Quantization Support FP8, INT4, AWQ FP8, INT8, BF16 FP8, BF16
Memory Footprint (INT4) 2.6 GB VRAM 14.8 GB VRAM 42.0 GB VRAM
Classification Accuracy (MMLU-Route) 92.4% 96.8% 88.1% 
Perplexity Multi-Stage Decision Flow
┌─────────────────────────────────────────────────────────────┐
│ Inbound Query / Document / Screenshot Stream                │
└──────────────────────────────┬──────────────────────────────┘
                               │
                               ▼
┌─────────────────────────────────────────────────────────────┐
│ Decider Flash v1.2 (4B - 14ms Latency)                      │
│ ├── Categorical Intent Classification                       │
│ └── Confidence Calibration Score [0.00 to 1.00]             │
└──────────────┬───────────────────────────────┬──────────────┘
               │ (Confidence >= 0.85)          │ (Confidence < 0.85)
               ▼                               ▼
┌─────────────────────────────┐ ┌─────────────────────────────┐
│ Immediate Execution Path    │ │ Decider v1.2 (27B - Deep)   │
│ - Direct API / Cache Hit    │ │ - Multi-Domain Disambig     │
└─────────────────────────────┘ └─────────────────────────────┘

Confidence Calibration Engineering

A central failure mode of standard language models is overconfidence on uncertain queries. Perplexity trained the Decider series using Expected Calibration Error (ECE) loss objectives alongside cross-entropy loss:

  1. True Probability Mapping: If Decider predicts an 85% likelihood that a prompt requires live financial data, empirical testing confirms the model is correct 85% of the time across test samples.
  2. Abstention Signaling: When incoming prompts contain irreconcilable contradictions, the model outputs an explicit ABSTAIN token rather than generating a speculative guess.
  3. Multimodal Grounding: The vision head processes UI screenshots, charts, and diagrams to classify whether visual elements require OCR extraction or layout analysis.

Community Adoption and Integration

Because Decider Flash v1.2 fits comfortably within 3 GB of VRAM, developers can embed the model directly onto client edge nodes, local gateway routers, and API reverse proxies to filter inbound traffic before paying for cloud inference.