Perplexity open-sources two Decider models with new text-locating capability
Perplexity released Decider v1.3 and Decider Flash v1.3 on November 12, 2026, with open weights.
· Last verified: 14 Nov 2026 (Perplexity, Unite.AI, AI Weekly, Hugging Face)

Summary
- Perplexity released Decider v1.3 and Decider Flash v1.3 on November 12, 2026, with open weights.
- The models return a decision or an exact text span instead of full language-model output.
- Developers get cheaper request triage and personal-data detection without running a general-purpose model.
The latest
Perplexity released two open-weight decision models, Decider v1.3 and Decider Flash v1.3, on November 12, 2026, both carrying a new capability called locate. Chief executive Aravind Srinivas said on X that the models are open source and free on OpenRouter for one week. With locate, a user describes what to find and the model returns the matching text plus its position.
Details
- The larger model: Decider v1.3 is built on Qwen3.8-27B and ships at roughly 49 gigabytes of weights in BF16 format, according to the model cards on Hugging Face. Self-hosting either model requires a CUDA graphics processor and working memory exceeding the size of the weights.
- The small model: Decider Flash v1.3 is built on Qwen3.5-4B at about 8.5 gigabytes, distilled from the larger model using KL divergence against the teacher model's output probabilities. The size gap is the main lever for developers weighing local deployment against hosted access.
- How locate works: The capability returns the exact matching text along with character start and end positions, and supports named-entity recognition, personal-information detection and extractive question answering. It returns an empty list when nothing matches, and is text-only.
- The constraints: Locate is capped at 128 questions per request. Requests mixing locate questions with decision questions require separate forward passes, which increases response latency — a cost developers batching large workloads would absorb directly.
- Classification scores: On query classification the larger model reaches 91.6% average accuracy across 63 tasks, a gain of 6.54 percentage points over Perplexity's production classifiers, with references relabeled by a frontier language model.
- Reranking scores: On reranking retrieval results the model posts an average nDCG@10 of 83.99 across 72 tasks, 1.81 points above Qwen3-Reranker-8B, according to the published benchmarks.
- Production agreement: On a deduplicated sample of one million API requests covering 2,318,887 questions, v1.3 agreed with frontier-model judges in 93.98% of cases, against 93.68% for v1.2. The report notes this measures agreement with model-generated references, not accuracy against ground truth.
- The cost figures: Running that sample cost $26.47 for v1.3 and $13.24 for Flash. No per-token price was announced, no local hardware cost was calculated, and no matched latency results were published alongside the accuracy numbers.
- Agent security logs: In a single day Perplexity processed one billion tokens across 7,000 sessions and the system proposed 14 rule fixes, six of which shipped within two days. Replaying a week of logs with the fixes cut the busiest rule's volume by 52%.
- Licensing: Weights are downloadable under Apache-2.0 on Hugging Face as perplexity-ai/pplx-decider-v1.3 and perplexity-ai/pplx-decider-flash-v1.3. Hosted access runs through OpenRouter and Perplexity's Decisions API. The company recommends calibrating decision thresholds on representative labeled examples.
Between the lines
The headline accuracy gain over v1.2 is 0.3 percentage points, and the company itself frames that benchmark as agreement with model-generated references rather than ground truth. The sharper selling point is structural: locate returns a character span, and Flash ran the same million-request sample at half the cost of the larger model.
What's next
The free OpenRouter window runs one week from November 12, after which hosted pricing becomes the practical test. Watch for a published per-token price and for matched latency figures covering mixed locate-and-decision requests.