Perplexity has released Decider v1.3, an open-weight decision model that outputs probabilities instead of text to classify queries, score answers, or locate specific text spans.
The release includes two variants: pplx-decider-v1.3 and pplx-decider-flash-v1.3. Both models are available on Hugging Face under an Apache 2.0 license, as well as via the Perplexity Decisions API and OpenRouter.
Model specs and hardware
The main model relies on a Qwen3.8-27B backbone, while the Flash variant uses Qwen3.5-4B. Hugging Face lists the parameters as 26B and 5B respectively, though Perplexity documentation cites 27B and 4B.
Running the full v1.3 weights requires a CUDA GPU capable of holding approximately 49 GiB of BF16 data. The Flash model needs about 8.5 GiB. Neither variant offers quantized versions.
Context windows support up to 256,000 tokens via the API. The default setting for local deployment is 32,768 tokens.
Architecture and training
Decider adapts a standard generative large language model into a classifier. The output layer is trimmed to specific option tokens, such as A, B, or C. A linear decision head reads the hidden state at a <|marker|> token to calculate probabilities for each choice.
Perplexity switched full-attention layers to bidirectional mode. Internal tests showed this change improves performance. The model also handles multiple questions within a single input differently than previous versions. In v1.2, every question could see every other question, which sometimes shifted answers. In v1.3, the shared state sees all questions, but individual questions only see the state and themselves. All questions regarding one input still process in a single forward pass.
Training occurred in three stages:
- A full fine-tune on hundreds of public datasets.
- A LoRA specialization adapter.
- A LoRA vision adapter.
Weights are averaged across three different seeds. The Flash model was trained by minimising KL divergence to match the outputs of the larger model.
The Locate feature
The new Locate function returns exact text spans with start and end character offsets. Users can request short labels like 'location' or specific instructions such as 'Extract the city name where the job is located.'
It targets named entity recognition, personally identifiable information detection, and extractive question answering. A new <|locate|> token is added for each requested field. A span head plus a LoRA adapter is trained for this task, while the backbone and decision head remain frozen. The adapter is kept unmerged because merging it hurt decision accuracy.
Benchmark results
All figures below are self-reported by Perplexity.
- Query classification: The model achieved 91.6% mean accuracy across 63 tasks, scoring 6.54 points higher than production classifiers.
- Domain classification: It outperformed Jev 1.14 in all seven tested domains, with a margin of up to 12.9 points in finance.
- Retrieval and reranking: It scored 83.99 mean nDCG@10 across 72 tasks. This is 1.81 points above Qwen3-Reranker-8B and 2.24 points above Jev 1.14.
- Multilingual: v1.3 scored 88.16% on 19 tasks across eight languages. Flash scored 86.50%, and Jev 1.14 scored 85.41%.
- Locate: The model averaged 69.0 on 17 out-of-domain tasks. Flash averaged 63.6. Competitors Liquid d1 and GLiNER large v2.5 scored lower at 43.3 and 33.6 respectively.
- Image classification: v1.3 reached 76.99% on five tasks, edging Jev 1.14 by 1.47 points. Flash scored 72.77%.
- Real-world requests: On a test of one million requests containing 2,318,887 questions, the model agreed with frontier LLM judges 93.98% of the time. Jev 1.14 agreed 93.35% of the time.
- Robustness: Answers flipped on only 0.27% of reorder or relabel transformations. Jev 1.14 saw flips between 2.69% and 4.84%.
Comparison table
| Feature | Decider v1.3 | Decider Flash v1.3 | Jev 1.14 | Qwen3-Reranker-8B |
|---|---|---|---|---|
| Size | 27B (docs) / 26B (HF) | 4B (docs) / 5B (HF) | Not disclosed | 8B |
| Context | API < 256K; local 32,768 | API < 256K; local 32,768 | 262,144 | 32K |
| License | Apache 2.0 | Apache 2.0 | Not disclosed | Apache 2.0 |
| Modality | Text, images, Locate | Text, images, Locate | Text, images | Text |
| Hardware | ~49 GiB BF16 | ~8.5 GiB BF16 | Hosted only | Not disclosed |
| Price | $0.02 / 1M input | $0.01 / 1M input | $0.042 / 1M input | Open weights, self-host |
| Retrieval nDCG@10 | 83.99 | Not disclosed | 81.75 | 82.18 |
| Multilingual accuracy | 88.16% | 86.50% | 85.41% | Not reported |
| Judge agreement | 93.98% | Not disclosed | 93.35% | Not reported |
| Cost for 1M requests | $26.47 | $13.24 | $68.10 | Not reported |
These benchmark rows are self-reported numbers from Perplexity’s launch post.
Deployment details
The API endpoint is POST https://api.perplexity.ai/v1/decisions. Model IDs are pplx-decider-v1.3 and pplx-decider-flash-v1.3, with -latest aliases available.
Pricing is $0.02 for v1.3 and $0.01 for Flash per one million input tokens. Output is free.
Latency tests showed an approximately 243K-token request took 16 seconds on v1.3 and 5 seconds on Flash. Locate requests allow up to 128 questions.
For local deployment, load the model using Decider.from_pretrained() from the pplx_decider package. Full-attention layers must run noncausally. Perplexity recommends v1.3 for nuanced rubrics and Flash for high-volume routing.
What it means
Developers gain a tool that separates decision logic from text generation. This reduces hallucination in classification tasks and allows for precise extraction of text spans without the model inventing details. The lower cost and open weights make it viable for internal routing systems or privacy-sensitive applications where data leaves the local environment.