Perplexity has released Decider v1.2, a 27-billion-parameter multimodal decision model, alongside Decider Flash v1.2, a smaller 4-billion-parameter version. The models are designed for structured decision-making tasks such as classification, content scoring, request routing and image-based evaluation, rather than conventional conversational text generation. Perplexity has made both sets of model weights available under the Apache 2.0 licence, giving developers greater flexibility to inspect, adapt and deploy the technology.
The release introduces a joint-request format that allows developers to ask several structured questions about the same input in one model pass. Both models accept text and images and return probabilities for predefined answers, including yes-or-no decisions, category selections and scores. Perplexity says the updated approach improves efficiency and evaluation performance compared with its earlier models. The release adds another option for developers building AI agents and automated workflows that need predictable, structured decisions instead of open-ended responses.
Key Takeaways
- Two models released: Decider v1.2 has 27 billion parameters, while Decider Flash v1.2 has 4 billion.
- Open-weight licensing: Both models use the Apache 2.0 licence.
- Multimodal input: The models process text and images, with image support up to 1.6 megapixels.
- Structured decisions: They answer yes-or-no, multiple-choice and scoring questions with probabilities.
- Joint requests: Developers can ask several questions about one input in a single model pass.
- Different priorities: The 27B model targets more nuanced decisions, while Flash is designed for lower latency and cost.
- Developer access: The models are available through Hugging Face, while Perplexity also provides access through its Decisions API.
What Are Perplexity Decider Models?
Most generative AI models are designed to produce text. A user submits a prompt, and the model generates a response token by token. That approach works well for writing, summarisation, explanations and open-ended conversations.
However, many business applications need a more constrained answer. A customer-support platform may need to decide whether a message is a billing complaint or a technical issue. A moderation system may need to estimate whether a post violates a particular policy. An AI agent may need to determine whether a browser page contains the information required to complete a task.
Using a general-purpose language model for these jobs can introduce unnecessary complexity. Developers must often instruct it to follow a fixed format, parse its response and handle cases where the output does not match the expected structure.
Perplexity’s Decider models are designed for these situations. Rather than generating a paragraph, they evaluate predefined questions and return probabilities for the possible answers.
For example, a developer could provide a customer message and ask:
- Is this message related to billing?
- Which department should receive it?
- How urgent is the issue on a scale of one to five?
The model can return structured results for those questions. An application can then use the scores to route the ticket, prioritise a response or send uncertain cases for human review.
This makes Decider a specialised component for AI applications rather than a direct replacement for a general-purpose chatbot.
Decider v1.2 vs Decider Flash v1.2
The two releases target different requirements. Decider v1.2 uses a larger model with 27 billion parameters, while Decider Flash v1.2 has 4 billion parameters.
A larger model can have more capacity for nuanced distinctions, but it generally requires more computing resources. A smaller model can be more practical for high-volume workloads where latency and operating costs matter.
| Feature | Decider v1.2 | Decider Flash v1.2 |
|---|---|---|
| Model size | 27B parameters | 4B parameters |
| Base model | Qwen3.8-27B | Qwen3.5-4B |
| Input types | Text and images | Text and images |
| Primary focus | More nuanced decisions | High-volume, lower-cost decisions |
| Output | Structured answers with probabilities | Structured answers with probabilities |
| Licence | Apache 2.0 | Apache 2.0 |
The Flash model is derived from the larger decision-model family through distillation. In this process, a smaller model learns from a larger model’s outputs and probability distributions, aiming to preserve useful behaviour with fewer parameters.
That does not mean the two models will perform identically. Perplexity’s published evaluations show that Flash retains much of the larger model’s performance, but the 27B model scores higher on the overall Decision Index and vision evaluation reported in its model card.
The appropriate choice will depend on the task. A system that processes thousands of straightforward support tickets may prioritise speed and cost, while a workflow requiring finer distinctions between similar categories may benefit from testing the larger model.
The Joint-Request Format Is the Main Technical Change
One of the most important changes in version 1.2 is how the models process multiple questions about the same input.
Earlier versions typically handled one question per forward pass. If an application needed to classify a document, determine its urgency and score its quality, it could require separate model passes for each question.
The new joint-request format places the input and multiple questions into one sequence. The model then produces a separate result for each question.
This can reduce repeated processing when several questions concern the same content. It also allows the model to evaluate related questions together, which may be useful when the questions describe different aspects of one decision.
Consider an AI system reviewing an uploaded invoice. It might need to determine whether the document is an invoice, identify the currency and estimate whether the total is consistent with the line items. Instead of repeatedly processing the same document, the application can group suitable questions into one request.
Perplexity says the joint approach can substantially improve throughput. Some secondary reporting cites speed improvements of up to five times, while Perplexity’s own technical materials report throughput gains that vary by workload and inference setup. These figures should not be interpreted as a universal speedup for every application.
Grouping questions also requires care. Related questions about a single document may benefit from joint processing, while questions comparing several competing candidates may be better handled separately. Developers should test both approaches on their own workloads.
Multimodal AI: Reading Text and Images Together
Both Decider models support images alongside text. This makes them useful for applications that need to interpret screenshots, documents, product photos or other visual information.
For example, an e-commerce platform could use a decision model to classify a product image, assess whether the photograph matches a listing category or identify which moderation workflow should review an upload.
A browser-based AI agent could also use a screenshot to determine whether a page displays a login form, a confirmation message or an error state. The result could help the agent decide which action to take next.
The models support images up to 1.6 megapixels, according to their model cards. This is sufficient for many common screenshots and document images, although the usefulness of the output depends on image quality, task complexity and the model’s learned capabilities.
Multimodal input is particularly relevant to AI agents because their environments are not always represented as clean text. Agents may encounter visual interfaces, charts, receipts, product images and other information that requires classification before the next step.
Decider does not independently complete an entire workflow simply because it can inspect images. It provides structured decisions that other software can use to determine what happens next.
What the Models Can Be Used For
The models are intended for tasks where an application needs a defined answer rather than a long explanation.
Customer-support routing
A support platform can classify incoming messages by topic, urgency or department. High-confidence decisions can be routed automatically, while uncertain cases can be sent to a human operator.
Content moderation
A platform can ask whether a post matches a specified policy category or whether an image requires further review. Such systems still need careful testing, clear policies and human oversight for consequential decisions.
AI-agent control
An agent can use a decision model to evaluate the current state of a task and determine which of several permitted actions is most appropriate. This can help separate decision-making from the software that executes actions.
Document and image assessment
Businesses can classify invoices, receipts, forms and screenshots, or score whether an uploaded document meets a predefined requirement.
Quality checks
Developers can assess whether a generated response meets a rubric, follows a required format or includes specified information. The output can support a quality-control pipeline, although automated scores should not be treated as infallible judgments.
These applications share a common requirement: they need decisions that software can consume consistently.
What Do the Benchmarks Show?
Perplexity’s published model cards report improvements over its earlier decision models on the Decision Index, a benchmark suite covering areas such as knowledge, language, retrieval and tool-related decisions. The company also reports results on a separate vision-focused evaluation.
| Published evaluation | Decider v1.2 (27B) | Decider Flash v1.2 (4B) |
|---|---|---|
| Overall Decision Index 0.3, joint requests | 68.09 | 58.88 |
| Overall Decision Index 0.3, one question per request | 68.09 | 60.05 |
| Vision Decision Index 0.3 | 77.43 | 72.53 |
Figures are from Perplexity’s published model cards. Benchmark versions and request formats affect the results, so comparisons should use matching evaluation settings.
The figures suggest that the larger model performs better overall on the reported evaluations, while Flash delivers a substantial portion of that performance with fewer parameters.
However, benchmark scores have limitations. A model can perform well on a particular evaluation and still struggle with unusual inputs, ambiguous instructions or categories that differ from its training data. The results are also reported by the model developer, so independent evaluations would provide additional evidence about how the models perform in different environments.
Businesses should test the models against their own data, including edge cases and examples where a mistaken decision would be costly. They should also measure latency, throughput, calibration and the frequency of errors rather than relying on one overall score.
Open-Weight Licensing and Developer Access
Perplexity has released the model weights under Apache 2.0, a permissive licence commonly used for open-source software and AI projects. The licence generally allows commercial use, modification and redistribution subject to its terms and applicable notices.
Open weights give developers more control over how a model is deployed. Depending on their infrastructure and technical requirements, teams can investigate self-hosting, adapt the model or integrate it into their own applications.
The models are available through their Hugging Face repositories:
perplexity-ai/pplx-decider-v1.2perplexity-ai/pplx-decider-flash-v1.2
Perplexity also provides a Decisions API for developers who prefer a managed service. The published API pricing is $0.02 per million input tokens for the 27B model and $0.01 per million input tokens for Flash, with output tokens listed as free under the announced pricing. Actual application costs can also include infrastructure, image processing, data preparation and engineering work.
Self-hosting is not automatically cheaper. The larger model requires suitable hardware and deployment expertise, while a managed API may be simpler for teams that do not want to operate their own inference infrastructure.
The Bigger Picture
Perplexity’s Decider release reflects a growing interest in specialised models that handle a narrow but important part of an AI workflow. Instead of using a general-purpose model for every task, developers can combine conversational models with smaller decision systems that classify inputs, estimate confidence and route work to the right component.
The open-weight release also gives developers more options for deployment and customisation. Its practical value will depend on how well the models perform on real-world data, how reliably their probabilities can guide decisions and whether the efficiency gains justify integrating a dedicated model into an existing system.
Looking Ahead
Developers will likely evaluate Decider v1.2 and Flash v1.2 against their existing classification and routing systems, paying particular attention to accuracy, speed and operating costs. The joint-request format may be especially useful in pipelines that repeatedly evaluate several related properties of the same document, message or image. Real-world testing will determine where the larger model’s additional capacity is worth the extra computing requirements.
More broadly, decision models could become a standard component of agentic AI systems, working alongside language models, search tools and execution environments. They will not eliminate the need for careful software design or human oversight, but they offer another way to make AI workflows more structured and predictable.
Frequently Asked Questions
1. What is Perplexity Decider v1.2?
It is a 27-billion-parameter multimodal decision model designed to classify, score and route information by answering predefined questions with structured probabilities.
2. What is Decider Flash v1.2?
It is the smaller 4-billion-parameter version, designed for applications where lower latency and cost are important while retaining much of the larger model’s reported benchmark performance.
3. Are the models open source?
Perplexity has released both models’ weights under the Apache 2.0 licence. Developers should review the licence and repository notices for the applicable terms.
4. Can Decider models process images?
Yes. Both models support text and images, with image input up to 1.6 megapixels according to their model cards.
5. How is a decision model different from a chatbot?
A chatbot generally generates conversational text, while a decision model evaluates predefined options and returns structured answers or probabilities that software can use directly.
6. How much does the Perplexity Decisions API cost?
Published input pricing is $0.02 per million tokens for Decider v1.2 and $0.01 per million tokens for Decider Flash v1.2, with output tokens listed as free. Developers should verify current pricing before deployment.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
