Learn

PLLM is a Rust-first, Python-friendly runtime and research harness for private large language model inference across multiple parties. Its goal is to let a client use remote compute without giving any one compute party the complete request, model state, and output.

PLLM combines two systems:

• The inference runtime assigns work to the client, preparation service, and inference service while preserving an explicit trust boundary.
• The research harness turns papers and hypotheses into independent Rust and Python implementations, then records provenance, plans, tests, benchmarks, limitations, and publication review separately.

PLLM is being built for high performance, but performance depends on the exact model, protocol, hardware, network, and workload. Privacy depends on the selected protocol, correct implementation, and stated non-collusion assumptions. Treat the project's economic and safety benefits as goals; use evidence records for measured results and current support for executable coverage.

Start here if you need to decide whether a PLLM workflow matches your privacy, performance, or deployment requirements:

1. Privacy and threat models
2. Parties and offline work
3. Masked linear inference
4. Arithmetic and Boolean garbling
5. Numeric semantics and model quality
6. Reading research and evidence

These pages explain the concepts. The reference section lists exact APIs and schemas. Current support reports which combinations are available. Integrations shows how the trusted local gateway serves OpenAI-compatible SDKs and coding agents.