Perplexity open-weight models target routing and search

The Hugging Face release pairs a 27-billion parameter decider with a contextual embedding model designed to lower infrastructure bills.

Perplexity AI published two Perplexity open-weight models on Hugging Face aimed at developer infrastructure: pplx-decider-v1-27b, designed for structured classification and workflow routing, and pplx-embed-v2-context-9b-preview, built for vector retrieval pipelines. Chief executive Aravind Srinivas announced the models alongside a commercial API intended to reduce operating costs for autonomous agent workflows.

Structured decisions instead of open-ended text

The pplx-decider-v1-27b model builds on Alibaba’s Qwen3.8-27B. It processes text and image inputs within a 250,000-token context window, measuring the discrete text segments an artificial intelligence reads at once. Its design diverges from standard chat models. Instead of drafting prose or engaging in conversations, it calculates probability distributions across predetermined choices. Typical workloads include sorting customer tickets to specific queues or returning binary yes-or-no determinations.

In evaluation figures published by Perplexity, the decider averaged 85.71% across 11 classification benchmarks, topping base Qwen (74.76%) and alternatives such as Jev (84.51%). On RAGTruth, a benchmark tracking factual consistency, it scored 88.80%, while its v1.1 release reached 94.5% on decision benchmarks. For teams avoiding self-hosted setups, the Decisions API launched at $0.04 per million input tokens with free outputs, later cut to $0.02. Deploying the model locally requires high-end hardware: inference calls for CUDA-capable GPUs and approximately 49 GiB of dedicated video memory.

Embeddings that keep the full document in view

The second model is pplx-embed-v2-context-9b-preview, which carries roughly 9 billion parameters. In retrieval-augmented generation (RAG) systems, long documents are converted into numerical representations known as vectors. Standard retrieval engines divide text into isolated chunks, which strips away the surrounding context of the full page. Perplexity’s embedding model encodes those fragments together, allowing each segment’s vector to retain the context of the entire source document.

Chief technology officer Denis Yarats noted that the model was trained through distillation from a query-aware context compression system. It produces non-normalized INT8 vectors across 2,048 dimensions, supporting Matryoshka dimension reduction down to 1,024 dimensions. At that lower dimension, each vector requires just 1 KB of storage, cutting memory footprints eightfold compared to standard Float32 representations. In blind evaluations run by vector search firm turbopuffer across nearly 39,000 documents, the model exceeded voyage-context-4 by 14.4 percentage points in Answer Recall@10.

Limits of Perplexity open-weight models and the cloud race

Releasing open weights is not the same as open-sourcing a complete project. Perplexity shared model weights under terms inherited from Qwen, but the underlying training datasets and proprietary curation code remain proprietary. In addition, the company explicitly labeled embed-v2 as a preview. Its architecture and vector dimensions could change before a stable release, meaning vector databases populated today might require complete re-indexing later.

The releases underline a deliberate shift toward enterprise infrastructure, putting Perplexity in direct competition with OpenAI, Anthropic, and Cohere. In multi-step agent systems, routing requests or retrieving text snippets through frontier models is an expensive design choice. Perplexity is betting that targeted, cost-efficient models can take over those routine technical steps for pennies.