pplx-decider v1.1 27B VRAM requirements

Which GPU holds the published checkpoint?

GiB means 1,073,741,824 bytes. File bytes include safetensors headers. The separate head is 2.49 MiB; it is included in the total, not an extra download after the estimate.

Why a chat-model VRAM estimate is insufficient

The checkpoint replaces the language-model output head with a 255 × 5,120 BF16 readout. Its full-attention layers read the input noncausally; the linear-attention layers retain their own behavior. There is no chat decoding loop to size with the site’s usual growing generation cache or tokens-per-second estimate.

Longer input, more images and a larger batch still increase activations and attention working memory. The supplied loader limits the complete request to 8,192 positions. For the publisher’s evaluation setting, cap each image at 512 × 512 pixels; its default higher cap permits up to 4,096 visual tokens per image and costs more memory.

INT4 and FP8 weight arithmetic alone cannot establish a working Decider deployment. A quantized export must preserve the readout, saved attention behavior and calibrated probabilities. We have not tested one.

Load the official decision implementation

Use Python 3.12 or newer and a CUDA environment with enough free memory. Download the repository, then install its pinned requirements.txt. The model card mentions authenticated access; check the repository’s current access requirements when downloading.

hf download perplexity-ai/pplx-decider-v1.1-27b --local-dir pplx-decider-v1.1-27b

pip install -r pplx-decider-v1.1-27b/requirements.txt

import sys
from pathlib import Path
from huggingface_hub import snapshot_download

checkpoint = Path(snapshot_download("perplexity-ai/pplx-decider-v1.1-27b"))
sys.path.insert(0, str(checkpoint / "source" / "src"))
from autojev.model import DecisionModel, answer

model = DecisionModel(checkpoint, device="cuda", max_image_pixels=512 * 512)
question = {
    "type": "choice",
    "instructions": "Route this request to a team.",
    "criteria": {"billing": "Payments and refunds", "support": "Integration faults"},
}
probabilities = model.predict([{"state": "My API request fails.", "question": question}])[0]
print(answer(question, probabilities))

Example follows the publisher’s DecisionModel API. It has not been executed on a GPU by ModelVRAM. Use predict’s calibrated distribution; do not apply the saved temperature a second time.

Sources and related decision models

Read on October 9, 2026: Perplexity’s model card, config.json, checkpoint files and decision_config.json. The weight total sums all 11 model shards and readout.safetensors.

Compare with LiquidAI d1-omni, the smaller text, image and audio decision model, or Laya for text-only decisions. How we distinguish estimates from measurements.