How to Run Perplexity's Decider v1.2 Locally: VRAM and Setup Guide
A practical guide to running pplx-decider-v1.2, Perplexity's 27B Qwen3-based decision model, covering VRAM needs and real-world performance.
What is Perplexity’s Decider v1.2?
Decider v1.2 is a decision model built by Perplexity on the Qwen3 27 billion parameter backbone. Unlike a general chat model, it’s purpose built to take a piece of input (text or an image) and return structured, scored answers to specific questions, like which team should handle a support ticket, whether a customer is blocked, or how frustrated someone sounds on a numeric scale. Running it locally requires roughly 28 to 29 GB of VRAM given its 27B parameter size, putting it in range of a single high-memory consumer or prosumer GPU rather than a multi-GPU setup.
TL;DR
- Decider v1.2 is a 27 billion parameter decision model from Perplexity, built on the Qwen3 27B backbone, designed to score and classify inputs rather than generate free-form chat responses.
- Running it locally needs approximately 28 to 29 GB of VRAM, which fits on a single 32GB or larger GPU without needing to split the model across multiple cards.
- The model processes a full request in one pass, meaning multiple questions about the same input get answered together instead of requiring separate calls.
- Version 1.2 adds image support up to 1.6 megapixels, extending the model beyond text-only decision tasks into basic visual reasoning.
- Perplexity’s own benchmark card shows the biggest gains over v1.1 came in language and knowledge tasks, while tool-use scoring stayed roughly flat between the two methods tested.
- Independent testing found the model strong at reading structured data from images (like spec sheets) but inconsistent on nuanced visual judgment and mixed-sentiment text analysis.
- Perplexity has acknowledged that some benchmark gains came from training on official benchmark splits, which is worth factoring in when judging reported scores.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
What hardware do you need to run Decider v1.2?
At 27 billion parameters, Decider v1.2 needs around 28 to 29 GB of VRAM to run. That number lines up with what you’d expect from a dense model of this size at standard precision, and it means the model fits comfortably on a single GPU with 32GB or more of memory. You won’t need a multi-GPU rig or enterprise-grade hardware to get it running, which is a meaningful difference from larger frontier models that require sharding across several cards.
If your GPU has less than 28GB of VRAM, the usual workarounds apply: quantization to reduce the memory footprint, offloading some layers to system RAM, or running on a cloud GPU instance sized for the mid-20GB to 30GB range. None of these are unique to Decider specifically, they’re standard techniques for any 20B+ parameter model, but they’re worth knowing going in since the model’s size sits right at the edge of what a single consumer GPU (like a 24GB card) can handle without adjustment.
How does Decider v1.2 differ from v1.1?
The core architectural change in v1.2 is that it reads the entire request in a single pass. Practically, this means if you ask several questions about the same message or image (which team should handle this, is the customer blocked, how frustrated are they), the model answers all of them together rather than requiring repeated calls with separate context windows. That’s a meaningful efficiency gain for any pipeline that needs multiple structured judgments from one input.
The other headline addition is image support. Decider v1.2 can now process images up to 1.6 megapixels alongside text, extending its decision-making role from pure text classification into basic visual tasks. Perplexity’s own model card reports that v1.2 outperforms v1.1 overall, with the largest improvements showing up in language and knowledge tasks. Tool-use scoring, which covers decisions about when and how to invoke external tools, came out close between the two approaches tested, suggesting that’s an area where the upgrade didn’t move the needle much. Image performance is also reported as improved over v1.1, though real-world testing suggests that claim deserves scrutiny (more on that below).
Is Decider v1.2 accurate enough for production use?
It depends heavily on the task. In testing by AI educator Fahad Mirza, who ran the model through a series of text and image based decision tasks, results were a mixed bag.
On straightforward text classification, the model performed well. Given a real customer support message describing a failed API integration (a 401 error after an API key rotation), Decider correctly identified the right team (support, at 99.8% confidence), correctly flagged the customer as blocked (98.4%), and returned a frustration score that reasonably matched the tone of a polite but stuck customer.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
It also did well on a task that played to a clear strength: reading structured data directly off an image. Given a GPU spec sheet and asked to identify which memory configurations the card came in, the model correctly picked both right answers with 99.7% confidence and put almost no weight on the wrong options. That’s a clean result on a task with a verifiable ground truth.
Where it struggled was anything requiring judgment layered on top of perception. Shown an AI-generated image of a passenger with a visibly incorrect seatbelt position, the model rated the seatbelt as “worn correctly” at 65.9%, a lean toward the wrong answer rather than a confident correct one. When the same kind of image was paired with a question about Indian road safety law specifically, the model again leaned toward “correct” at 89.2% despite the belt sitting under the shoulder rather than over it, the kind of detail a 27B vision-capable model should catch. It also misjudged a customer message using negation and idiom (“I can’t say I’m not disappointed”), scoring satisfaction at 11.2% when the sentiment was genuinely mixed rather than clearly negative.
The pattern that emerges: Decider v1.2 is dependable for tasks with a clear, checkable answer (classification, routing, reading explicit data), but less reliable when a task requires combining visual perception with external rules or reasoning about ambiguous, mixed-signal language.
Should you train or fine-tune expectations around benchmark scores?
Worth noting: Perplexity has stated that some of Decider v1.2’s benchmark gains came from training on official benchmark splits. That’s a detail worth keeping in mind when comparing the model’s reported scores to other decision models, since training directly on benchmark data tends to inflate performance on that specific benchmark relative to real-world generalization. It doesn’t mean the model is bad, but it does mean the gap between benchmark numbers and actual task performance (like the seatbelt and sentiment examples above) may be larger than the scorecard suggests.
How does Decider v1.2 compare to other decision models?
Decider v1.2 competes in a growing category of small-to-mid sized models built specifically for structured decision tasks rather than open-ended generation. Testing across this category has found that some sub-billion parameter models perform competitively on text classification tasks, and some vision-specific decision models handle image judgment tasks more reliably than Decider does here. At 27B parameters, Decider sits on the larger end of this category, which raises the bar for what users reasonably expect from it on visual reasoning tasks. The spec-sheet reading result shows it can hit that bar when the task is about extracting explicit information. The seatbelt and legal-rule tests suggest it doesn’t yet hit that bar when the task requires applying external knowledge or rules to what it sees.
Frequently Asked Questions
How much VRAM does Decider v1.2 need?
Approximately 28 to 29 GB, consistent with running a 27 billion parameter model at standard precision. A single GPU with 32GB or more of VRAM should run it without needing to split the model across multiple cards.
What backbone architecture does Decider v1.2 use?
It’s built on Qwen3’s 27 billion parameter backbone, adapted and trained by Perplexity specifically for structured decision and classification tasks rather than general chat.
Can Decider v1.2 process images?
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
Yes, as of v1.2 it supports images up to 1.6 megapixels in addition to text. It performs well on tasks like reading explicit data off a document or image, but has shown weaker performance on tasks requiring visual judgment combined with external rules or context.
Does Decider v1.2 support multiple languages?
It’s based on Qwen3, which has multilingual training, and in testing it correctly identified text written in German. However, it’s unclear from available testing whether it reliably comprehends non-English text at the same level as English.
Is Decider v1.2 good enough for production support workflows?
It performed well on routing and triage tasks involving clear-cut classification, such as identifying the right support team and detecting whether a customer is blocked. It performed less reliably on nuanced sentiment analysis and image-based judgment calls, so teams should validate it against their specific use case before relying on it unsupervised.
