A hands-on test of Perplexity's pplx-decider-v1.2 decision model on routing, sentiment, and vision tasks, including VRAM needs.
What is pplx-decider-v1.2? #
pplx-decider-v1.2 is a decision model from Perplexity AI, built on the Qwen3.8-27B backbone, designed to answer structured questions about text and images rather than generate open-ended chat responses. Instead of writing paragraphs, it returns confidence-scored answers to specific questions like “which team should handle this ticket” or “is this seatbelt worn correctly,” making it suited for routing, classification, and sentiment scoring pipelines. It is listed on Hugging Face under perplexity-ai/pplx-decider-v1.2, released with Apache 2.0 licensing, and is a 27 billion parameter model that needs roughly 28 to 29 GB of VRAM to run.
TL;DR #
- pplx-decider-v1.2 is a 27B parameter decision model on the Qwen3.8-27B backbone that scores answers to specific questions instead of generating free text.
- The model processes a whole request in a single pass , so multiple questions about the same message get answered together rather than one at a time.
- Vision support now handles images up to 1.6 megapixels, a new capability over the prior v1.1 release.
- In hands-on testing, the model handled routing and frustration scoring on a sample customer support ticket with plausible, actionable confidence scores.
- It struggled with seatbelt safety checks in two test images, including one where it missed an obvious violation under Indian road rules.
- It performed well reading structured data straight off an image , correctly picking GPU memory specs from a spec sheet with 99.7% confidence.
- A sentiment analysis test with mixed language returned a lopsided score that didn’t reflect the ambiguity in the text, suggesting a tendency to overweight negative words.
Remy is new. The platform isn't. #
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
How does the Decider model actually work? #
Unlike a typical LLM that produces a text response, pplx-decider-v1.2 is set up as a text-classification model (per its Hugging Face pipeline tag) that takes a request and a set of possible questions, then returns calibrated confidence percentages for each answer option. In a support routing example, feeding the model a customer message about a failing API integration produced three outputs in one pass: a team assignment (support, at 99.8% confidence), a blocked status (yes, at 98.4%), and a frustration score (1.28 on a four-point scale). That single-pass behavior is the core architectural change from v1.1, where multiple questions about the same input previously required separate calls. For teams building ticket triage or workflow routing on top of this, that’s a meaningful latency and cost improvement if the multi-question accuracy holds up.
What changed from v1.1 to v1.2? #
According to the model’s documentation, three things moved in this release. First, joint reasoning across multiple questions in a single pass, replacing what was likely a more sequential setup. Second, native image support up to 1.6 megapixels, extending the model beyond pure text classification into multimodal decision-making. Third, Perplexity’s own benchmark comparisons show the new version ahead of v1.1 overall, with the largest jumps specifically on language and knowledge tasks. Tool-use decisions, by contrast, showed the two versions scoring close to each other, meaning that particular capability didn’t see the same leap. Image understanding is listed as an area where v1.2 leads its predecessor, though real-world testing suggests that lead doesn’t necessarily translate into reliable accuracy (more on that below). It’s also worth noting Perplexity states some of the reported gains came from training on official benchmark splits, which is useful context when weighing how much the benchmark numbers generalize to novel inputs.
How did it perform on vision tasks? #
This is where the model’s limits showed up clearly in testing. Given an AI-generated image of a passenger with a seatbelt worn incorrectly, and asked a simple yes/no question about whether it was worn correctly, the model returned 65.9% confidence toward “yes,” a weak and arguably wrong lean on an image where the problem looked obvious to a human eye.
A second test pushed further by asking the model to judge seatbelt placement against Indian road safety law, a task requiring both visual recognition and rule application. The model answered “worn correctly” at 89.2% confidence, missing that the belt sat under the shoulder rather than over it, which fails the described legal standard. For a 27B parameter model marketed with upgraded image capability, that’s a notable miss, especially since it offered no indication it was reasoning about a specific legal rule at all, just a generic visual read.
Other agents ship a demo. Remy ships an app. #
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
The model did better on a more mechanical vision task: reading a real GPU data sheet image and picking the correct memory configurations from four options. Here it returned 99.7% confidence on the two correct answers and nearly zero on the incorrect ones, a clean result on a task that’s closer to OCR-plus-matching than open-ended visual judgment. It also correctly identified a German-language paragraph from an image as German, though it gave no indication of understanding the content beyond language recognition.
The pattern across these tests: the model handles extraction and lookup tasks (reading text or specs off an image) well, but struggles when a task requires applying contextual or legal judgment to what it sees.
Is the sentiment scoring reliable? #
Not consistently, based on testing with intentionally ambiguous text. A sample customer message using double negatives and an idiom (something like “I can’t say I’m not disappointed”) was fed to the model with a question asking whether the customer was satisfied overall. A fair human read of that message would land near the middle, since the phrasing is deliberately mixed. Instead, the model returned 11.2% confidence toward satisfaction, a clear skew toward “not satisfied” that overcorrected on the negative language in the sentence rather than capturing the actual ambiguity. That’s a meaningful failure mode for anyone planning to use this model for nuanced sentiment grading in production, since idiomatic or backhanded phrasing is common in real customer feedback.
By contrast, the routing and blocked-status outputs from the first support ticket test looked solid: the team assignment, blocked flag, and frustration score all landed in ranges that matched a reasonable human judgment of the same message.
Is pplx-decider-v1.2 worth using? #
It depends heavily on the task. For straightforward routing decisions and structured data extraction from images (reading specs, recognizing a language, matching known fields), the model performed well and returned confidence scores tight enough to act on without manual review. For tasks requiring contextual judgment, like applying a safety standard or reading mixed sentiment correctly, it fell short, including one case where it gave a high-confidence wrong answer on an obvious seatbelt violation.
Running it also isn’t trivial: at 27B parameters, it needs around 28 to 29 GB of VRAM, putting it out of reach of most consumer GPUs and into workstation or cloud territory. Given that smaller decision models (some under a billion parameters) have performed comparably or better on similar routing and vision tasks in independent testing, the VRAM cost here is a real consideration, not just a footnote. For teams evaluating decision models for production pipelines, it’s worth benchmarking pplx-decider-v1.2 directly against lighter alternatives on your specific task mix before committing resources to it.
Frequently Asked Questions #
What is pplx-decider-v1.2 used for?
It’s designed for decision tasks like routing customer messages to the right team, scoring sentiment or frustration levels, and answering structured yes/no or multiple-choice questions about text and images, rather than generating open-ended responses.
What hardware do you need to run it?
As a 27 billion parameter model, it requires approximately 28 to 29 GB of VRAM, which generally means a workstation-class or data-center GPU rather than a typical consumer card.
Can pplx-decider-v1.2 process images?
Yes. Version 1.2 added multimodal support for images up to 1.6 megapixels, a new capability compared to v1.1, which was text-focused.
Is it better than v1.1?
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Perplexity’s own benchmark comparisons show v1.2 ahead overall, with the biggest gains in language and knowledge tasks and image understanding. Tool-use decision accuracy, however, scored close to v1.1, showing little improvement there.
Is it reliable for sentiment analysis?
Testing found it can skew heavily toward negative readings on ambiguous or idiomatic text, returning lopsided confidence scores where a more balanced score would be expected. It’s worth double-checking sentiment outputs rather than trusting them blindly in production.