Perplexity releases Decider 27B weights for AI classification

Perplexity has published weights for its Decider 27B model on Hugging Face. The repository release record dates the commit to October 1, 2026. The weights are publicly accessible now, although the commit timestamp alone does not establish when the repository became public.

The model is also available through Perplexity’s Decisions API, priced at $0.04 per million input tokens. Output tokens are free, with no separate request charge. The service answers classification questions about text and images using probabilities.

Probability outputs for software

Developers can ask whether a statement is true, choose among supplied options or score content against an ordered rubric. The model returns numbers for software to act on and does not generate explanations. That makes routing requests and applying classification thresholds central uses.

The API reference permits up to 128 named questions about the same content in a single request. Total input must stay below 262,144 tokens, including the questions, and the request body must remain under 32 MiB. Responses preserve each question’s name and include token usage.

Current documentation sets a limit of 10 requests per second for each organization. A separate token limit applies to large bursts. For choice and score answers, the confidence field is the model’s certainty estimate and can differ from the probability assigned to the leading option.

Local deployment needs GPU capacity

The model card identifies Decider as a fine tune of Qwen3.8 27B and lists Apache 2.0 licensing. Its inference example requires Python 3.12 or later and a CUDA GPU with space for approximately 49 GiB of weights plus working memory.

Perplexity reports mixed results across individual benchmarks. Decider scores 88.80% on RAGTruth against 61.53% for the Qwen base model. On WinoGrande, its 83.30% trails the 90.70% reported for Jev. The Decider results were measured through Perplexity’s API, so these are vendor results rather than an independent assessment of local deployment.

Thresholds still need testing

Perplexity’s support ticket example shows why the surrounding rules matter. It checks serious impact before uncertainty, sending potentially critical tickets toward escalation even when their intent is ambiguous. The example also records small probability changes across repeated runs that could change a routing decision when a cutoff sits too close to the observed values.

For teams considering the model, the practical next step is to test representative cases and examine which mistakes cross an escalation threshold. The cookbook recommends labeling a few hundred tickets and comparing predicted routes with those labels before adjusting the cutoffs.

Our coverage of Cloudflare Clef and Clef Flash and Strands Decider 2B explores other decision models with different sizes and deployment limits.

Original AI generated conceptual artwork showing open model weights and branching decisions