Perplexity's Open-Source Surge: A Roundup of 6 AI Models and Tools

Perplexity has recently made a concentrated push in the open-source community, releasing a batch of projects spanning multimodal decision models, contextual embeddings, a local inference engine, privacy protection tools, a research agent benchmark, and an endpoint security solution. These projects cover everything from low-level inference performance to on-device privacy and enterprise security — a clear signal that Perplexity is much more than a search engine.

Two Core Models Claiming SoTA

The most attention-grabbing releases this round are two models that Perplexity claims have reached state-of-the-art (SoTA) performance.

pplx-decider-v1-27b is a 27-billion-parameter multimodal decision model. According to official figures, it achieves an average score of 85.7% across 11 benchmarks, outperforming the comparison target Jev. The "decision model" framing is noteworthy — this isn't a pure conversational or generative model, but one optimized for "making judgments and choices given available information." That kind of capability is a critical building block for agentic (Agent) systems.

pplx-embed-v2-context-9b-preview is a 9-billion-parameter contextual embedding model that achieves top performance on both the ConTEB and turbopuffer context-bench benchmarks. Unlike traditional static embeddings, contextual embeddings dynamically adjust vector representations based on the broader document context, directly improving retrieval quality in Retrieval-Augmented Generation (RAG) systems. For a company whose core strength is search, shipping a high-quality embedding model is a natural fit.

SoTA (State of the Art) refers to the best currently known performance on a specific task or benchmark. SoTA claims in AI should be treated with caution: different organizations use different comparison benchmarks, test set splits, and scoring methods, and self-reported numbers have limited value before independent third-party replication. Benchmarks themselves also carry a risk of "overfitting" — a model may have been specifically optimized for a particular benchmark suite and may not perform as well in real-world scenarios as its ranking suggests. SoTA is therefore more accurately read as "current best under these benchmark conditions," not as an absolute measure of capability.

Lily: A Local Inference Engine Built for Apple Silicon

Lily is a local inference engine purpose-built for Apple chips, and its technical approach is decidedly aggressive: it's implemented in Rust with custom Metal kernels, with no dependency on PyTorch or MLX — the two dominant mainstream frameworks.

Official benchmarks show that on the M5 Max chip, Lily's prefill speed is 1.23× faster than MLX-LM, and decode speed is 1.35× faster. Abandoning mature frameworks and rewriting from the kernel level up is engineering-intensive, but it also means lower overhead and tighter hardware adaptation. The emergence of purpose-built engines like this reflects a broader shift in on-device inference: from "it runs" to "it runs faster and leaner."

Prefill and decode are two fundamentally different phases in an inference engine. The prefill phase processes all input tokens at once, computing and caching the KV (Key-Value) attention matrices — it's a highly parallelizable matrix operation bottlenecked by compute. The decode phase generates output tokens one at a time; each step reads from the cached KV matrices and runs a single forward pass, making it bottlenecked by memory bandwidth. Apple Silicon's unified memory architecture (where CPU and GPU share the same memory pool) provides a natural advantage for local inference, but PyTorch and MLX's general-purpose abstraction layers introduce additional overhead. By writing Metal kernels directly, Lily bypasses that abstraction layer entirely, enabling fine-grained tuning for Apple GPU compute unit layouts and memory access patterns.

Two On-Device Privacy and Research Tools

PII-Tracer is a 600-million-parameter on-device PII (Personally Identifiable Information) classifier. Its core purpose is to determine whether a mixed-compute task should stay on the local Mac rather than be sent to the cloud. Perplexity claims it outperforms OpenAI's Privacy Filter on all 5 public benchmarks. Alongside the model, Perplexity also released the PII-TRACE benchmark, containing 13,000 conversations across 13 languages — multilingual coverage is a key highlight here, since the challenge of privacy detection lies precisely in how personal information manifests differently across languages and cultures.

WANDR is an evaluation benchmark for wide and deep research agents, comprising 500 tasks, each requiring 170,000 sourced records. That scale makes clear it targets genuinely complex, deep research scenarios — not simple Q&A — providing a far more demanding yardstick for evaluating research-oriented agents.

RAG (Retrieval-Augmented Generation) is the dominant approach for extending AI knowledge: before generating a response, the system retrieves relevant document chunks from an external knowledge base and incorporates them into the context for the language model to reference. Embedding models play a central role in RAG as the "indexing and retrieval" layer — they convert text into high-dimensional vectors and use vector similarity for semantic search. Traditional static embeddings (like word2vec) always produce the same vector for a given word; contextual embeddings dynamically generate vectors based on the full sentence or paragraph context, enabling more accurate capture of semantic nuance and reducing the likelihood of retrieving irrelevant documents — directly improving RAG answer quality. MCP (Model Context Protocol) is an open protocol proposed by Anthropic and increasingly adopted by multiple AI toolchains, defining how AI models interact with external tools and data sources. As more IDE plugins and browser extensions connect to AI capabilities via MCP, MCP configuration files themselves become a potential supply chain attack vector — malicious configs can redirect data flows or inject instructions, which is exactly why Bumblebee includes them in its scan scope.

Security Tools for Developer Machines

The final two tools turn their attention to developer and enterprise endpoint security.

Numbat is an agent detection and response tool for laptops and workstations. It ships with 52 built-in rules as a single Go binary and supports macOS, Linux, and Windows simultaneously. The single-file, cross-platform design significantly lowers the deployment barrier.

Bumblebee is a read-only supply chain scanner designed for developer machines, covering packages, MCP configurations, and editor and browser extensions. The inclusion of MCP (Model Context Protocol) configurations in its scan scope signals that Perplexity has already recognized that as AI toolchains proliferate, MCP is becoming a new potential attack surface.

A Deliberate Open-Source Strategy

Taken together, Perplexity's intent across these six projects is quite clear: it is building a complete open-source infrastructure stack around "AI-driven retrieval and agentic systems" — from core capabilities like decision models and embedding models, to the performance foundation of a local inference engine, to supporting tools for privacy, research evaluation, and endpoint security.

Perplexity has also teased that "more open-source contributions are coming." For the developer community, a company that started with a commercial search product being willing to continuously open-source models and tools — whether for technical brand building or ecosystem reasons — delivers genuinely usable resources regardless of motivation. One caveat worth noting: all performance figures cited above come from Perplexity's own official releases and have yet to be independently reproduced and verified by third parties.