pplx-decider

Perplexity's open decision models on Qwen3.8-27B and a 4B Flash: typed questions about a state, a probability for every option. Apache 2.0.

In short. pplx-decider is a family from Perplexity that answers questions about a message, a document or an image with a probability for each possible answer, instead of writing text. You can ask several questions about the same support ticket at once, such as which team should handle it and whether it is urgent, and get a number for every option in one pass.

Summary

Perplexity published its pplx-decider decision models on Hugging Face in October 2026 under Apache 2.0, with a 27B version on 1 October, v1.1 on 5 October, and on 15 October v1.2 together with a 4B Flash version. They are built on Qwen3.8-27B and Qwen3.5-4B. A request gives a state, which is text and optionally images of up to 1.6 megapixels, plus any number of typed questions. A question can ask for a choice among 1 to 255 options, a yes or no, or a score, and the model returns a probability for every option with no generated text. v1.2 reads the state and all questions in one sequence, with a readout at a marker after each question's options, so extra questions about the same state cost little and can inform each other, while v1.1 answered one question per forward pass. Flash is distilled from the 27B models on their training requests.

The numbers are Perplexity's own, on the public part of the Decision Index 0.3. v1.2 scores 68.09 with joint requests and 67.75 with one question per request, against 62.30 for v1.1, and 77.43 on the vision suite against 73.22. Flash scores 58.88, or 60.05 with one question per request, and 72.53 on vision. The card says some of the gain comes from training on the official training splits that several benchmarks ship, decontaminated against the test sets, and that both v1.2 checkpoints average the weights of several training runs. On decision data held out from training the probabilities stay somewhat over-confident, with an expected calibration error of about 0.05, against 0.014 on its development set. Grouping questions helps when they are facets of one decision and hurts when they are competing candidates, so the card advises one candidate per request when in doubt.

The 27B checkpoints are 52.2 GB in bf16 and Flash is 9.1 GB. Flash also has a Q8_0 GGUF of 5.16 GB plus a 0.67 GB vision projector, although stock llama.cpp server does not implement the checkpoint's protocol and a compatible runtime is needed. The Python package targets CUDA GPUs and has a fast path with torch.compile and CUDA graphs, and the card measures a median of 96 ms for one Decision Index request on it, against 131 ms for the reference path. Positions go up to 262,144 tokens, but training used sequences of up to about 32K tokens, and a single-fact retrieval test held to about 243K tokens, using roughly 95 GiB of GPU memory. In the 30 days to 11 November 2026, Hugging Face counted about 5,550 downloads of v1.2, 6,800 of Flash and 10,700 of the Flash GGUF.

Good for

  • Routing, triage and moderation steps that need several typed answers about one message in a single pass
  • Relevance scoring and reranking with yes or no questions, including images up to 1.6 megapixels
  • Latency-sensitive decisions with the 4B Flash, which also comes as a GGUF
  • Apache 2.0 weights from a well-known lab, with a documented Python package and request schema

Watch out for

  • The scores are Perplexity's own, and v1.2 trained on the official training splits of several benchmarks, so part of the gain on them reflects in-domain training
  • The checkpoint needs noncausal attention and a per-question readout, so ordinary inference servers cannot run it, and v1.1-style prompts do not work with it
  • Training used sequences up to about 32K tokens, and longer inputs were only spot-checked
  • Probabilities are somewhat over-confident on held-out tasks, so test them on your own labeled examples
  • The 27B weights alone need about 52 GiB of GPU memory, and the package targets CUDA GPUs

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the pplx-decider flash v1.2 (4B) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
pplx-decider flash v1.2 (4B) BF16 4.54B 10 GB 5.22 GB 3.13 GB Laptop or phone
pplx-decider flash v1.2 GGUF (Q8_0) INT8 4.54B 10 GB 5.22 GB 3.13 GB Laptop or phone
pplx-decider v1.2 (27B) BF16 26.1B 60 GB 30 GB 18 GB Gaming GPU
pplx-decider v1.1 (27B, one question per pass) BF16 26.1B 60 GB 30 GB 18 GB Gaming GPU

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is pplx-decider?
pplx-decider is a family from Perplexity that answers questions about a message, a document or an image with a probability for each possible answer, instead of writing text. You can ask several questions about the same support ticket at once, such as which team should handle it and whether it is urgent, and get a number for every option in one pass.
Can I use pplx-decider commercially?
Yes. pplx-decider is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is pplx-decider open source?
pplx-decider is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run pplx-decider on my own computer?
Yes. The smallest variant, pplx-decider flash v1.2 (4B), needs about 3.13 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was pplx-decider released?
pplx-decider was released by Perplexity on October 1, 2026.
How long a context window does pplx-decider have?
pplx-decider supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
AutoTrust Decider AutoTrust AI Lab 8.95B – 36B 3B active 256K Apache 2.0 (Permissive) 6.18 GB Laptop or phone
Clef Cloudflare 9.41B – 27.4B 64K Apache 2.0 (Permissive) 6.49 GB Laptop or phone
Liquid d1 Liquid AI 587M – 3.12B 128K LFM Open 1.0 (Conditional) 0.405 GB Laptop or phone
Qwen3.8-27B Alibaba Qwen 27.8B 256K Apache 2.0 (Permissive) 19 GB Gaming GPU