# PLLM

Release: 0.1.0

Private LLM inference documentation. Public weights; provider follows the protocol.

- [PLLM — Private LLM Inference](https://pllm.run/) ([Markdown](https://pllm.run/index.md)): Private language model inference. Client and server guides, the protocol, and reproducible research.
- [Get started](https://pllm.run/learn/start/) ([Markdown](https://pllm.run/learn/start.md)): Install PLLM, inspect a private inference plan, and run the local development benchmark.
- [Installation](https://pllm.run/learn/start/installation/) ([Markdown](https://pllm.run/learn/start/installation.md)): Install the Python package and verify the CLI without loading a model or contacting a service.
- [Inspect a private inference plan](https://pllm.run/learn/start/first-private-request/) ([Markdown](https://pllm.run/learn/start/first-private-request.md)): Define a reproducible experiment, lower a semantic model plan, and check its execution coverage.
- [Configuration](https://pllm.run/sdk/configuration/) ([Markdown](https://pllm.run/sdk/configuration.md)): Load, inspect, and export a reproducible PLLM experiment configuration.
- [Model plans](https://pllm.run/sdk/plans/) ([Markdown](https://pllm.run/sdk/plans.md)): Lower a model configuration, apply a compiler component, and inspect execution coverage.
- [Components](https://pllm.run/sdk/components/) ([Markdown](https://pllm.run/sdk/components.md)): Inspect versioned capabilities and apply compatible components to a model plan.
- [Build](https://pllm.run/sdk/build/) ([Markdown](https://pllm.run/sdk/build.md)): Define a reproducible PLLM system from configuration, models, components, and research methods.
- [Models](https://pllm.run/sdk/build/model-adapters/) ([Markdown](https://pllm.run/sdk/build/model-adapters.md)): Turn model-family configuration into a shared decoder plan.
- [Research methods and composition](https://pllm.run/research/recipes/build-method/) ([Markdown](https://pllm.run/research/recipes/build-method.md)): Turn a published method into a versioned component and a recorded plan change.
- [Understand](https://pllm.run/learn/understand/) ([Markdown](https://pllm.run/learn/understand.md)): Understand PLLM's architecture, trust boundaries, privacy assumptions, and evidence model.
- [Architecture](https://pllm.run/learn/understand/architecture/) ([Markdown](https://pllm.run/learn/understand/architecture.md)): Understand PLLM's plans, runtime stages, role ownership, and side effects.
- [Trust boundary](https://pllm.run/learn/understand/trust-boundary/) ([Markdown](https://pllm.run/learn/understand/trust-boundary.md)): Understand what the client, preparation service, inference service, and operators can access.
- [Privacy and assurance semantics](https://pllm.run/learn/understand/privacy-assurance/) ([Markdown](https://pllm.run/learn/understand/privacy-assurance.md)): State a specific privacy claim and interpret the evidence that applies to it.
- [Evidence and claim taxonomy](https://pllm.run/learn/understand/evidence-claims/) ([Markdown](https://pllm.run/learn/understand/evidence-claims.md)): Separate a design contract, an implementation, its evidence, and a publication claim.
- [Research APIs](https://pllm.run/sdk/research/) ([Markdown](https://pllm.run/sdk/research.md)): Use the Python SDK to benchmark plans, run scoped assurance checks, and produce immutable evidence records.
- [Benchmark](https://pllm.run/sdk/research/benchmark/) ([Markdown](https://pllm.run/sdk/research/benchmark.md)): Run scoped Python benchmarks and understand what the development dashboard does not measure.
- [Compare](https://pllm.run/research/recipes/compare/) ([Markdown](https://pllm.run/research/recipes/compare.md)): Compare runs with matching inputs and report every relevant cost and uncertainty.
- [Reproduce](https://pllm.run/research/recipes/reproduce/) ([Markdown](https://pllm.run/research/recipes/reproduce.md)): Compare an independent PLLM implementation with the published method it follows.
- [Assure](https://pllm.run/sdk/research/assure/) ([Markdown](https://pllm.run/sdk/research/assure.md)): Run a defined assurance check and report exactly what its result supports.
- [Operate](https://pllm.run/sdk/operate/) ([Markdown](https://pllm.run/sdk/operate.md)): Assign client, preparation, and inference responsibilities without weakening the privacy design.
- [Client boundary](https://pllm.run/sdk/operate/client-boundary/) ([Markdown](https://pllm.run/sdk/operate/client-boundary.md)): Understand which data and runtime state must stay with the client.
- [Provider roles](https://pllm.run/sdk/operate/provider-roles/) ([Markdown](https://pllm.run/sdk/operate/provider-roles.md)): Understand the responsibilities of deployment operators, preparation services, and inference services.
- [Deployment](https://pllm.run/sdk/operate/deployment/status/) ([Markdown](https://pllm.run/sdk/operate/deployment/status.md)): Understand current deployment support, local role orchestration, and production gaps.
- [Reference](https://pllm.run/sdk/reference/) ([Markdown](https://pllm.run/sdk/reference.md)): Look up exact CLI commands, Python APIs, schemas, components, and current support.
- [CLI reference](https://pllm.run/cli/reference/) ([Markdown](https://pllm.run/cli/reference.md)): Exact commands, arguments, and help text from the installed PLLM CLI.
- [pllm config](https://pllm.run/cli/reference/config/) ([Markdown](https://pllm.run/cli/reference/config.md)): Exact pllm config help from the installed PLLM CLI.
- [pllm components](https://pllm.run/cli/reference/components/) ([Markdown](https://pllm.run/cli/reference/components.md)): Exact pllm components help from the installed PLLM CLI.
- [pllm benchmark](https://pllm.run/cli/reference/benchmark/) ([Markdown](https://pllm.run/cli/reference/benchmark.md)): Exact pllm benchmark help from the installed PLLM CLI.
- [pllm dev](https://pllm.run/cli/reference/dev/) ([Markdown](https://pllm.run/cli/reference/dev.md)): Exact pllm dev help from the installed PLLM CLI.
- [pllm config show](https://pllm.run/cli/reference/config/show/) ([Markdown](https://pllm.run/cli/reference/config/show.md)): Exact pllm config show help from the installed PLLM CLI.
- [pllm config export](https://pllm.run/cli/reference/config/export/) ([Markdown](https://pllm.run/cli/reference/config/export.md)): Exact pllm config export help from the installed PLLM CLI.
- [pllm components list](https://pllm.run/cli/reference/components/list/) ([Markdown](https://pllm.run/cli/reference/components/list.md)): Exact pllm components list help from the installed PLLM CLI.
- [pllm components show](https://pllm.run/cli/reference/components/show/) ([Markdown](https://pllm.run/cli/reference/components/show.md)): Exact pllm components show help from the installed PLLM CLI.
- [pllm benchmark run](https://pllm.run/cli/reference/benchmark/run/) ([Markdown](https://pllm.run/cli/reference/benchmark/run.md)): Exact pllm benchmark run help from the installed PLLM CLI.
- [pllm gateway](https://pllm.run/cli/reference/gateway/) ([Markdown](https://pllm.run/cli/reference/gateway.md)): Exact pllm gateway help from the installed PLLM CLI.
- [pllm serve](https://pllm.run/cli/reference/serve/) ([Markdown](https://pllm.run/cli/reference/serve.md)): Exact pllm serve help from the installed PLLM CLI.
- [pllm serve inference](https://pllm.run/cli/reference/serve/inference/) ([Markdown](https://pllm.run/cli/reference/serve/inference.md)): Exact pllm serve inference help from the installed PLLM CLI.
- [pllm serve preparation](https://pllm.run/cli/reference/serve/preparation/) ([Markdown](https://pllm.run/cli/reference/serve/preparation.md)): Exact pllm serve preparation help from the installed PLLM CLI.
- [pllm dev dashboard](https://pllm.run/cli/reference/dev/dashboard/) ([Markdown](https://pllm.run/cli/reference/dev/dashboard.md)): Exact pllm dev dashboard help from the installed PLLM CLI.
- [Local experiments](https://pllm.run/cli/reference/gateway/local-experiments/) ([Markdown](https://pllm.run/cli/reference/gateway/local-experiments.md)): Serve declarative experiments, Python objects, and Python factories locally.
- [Provider connections](https://pllm.run/cli/reference/gateway/provider-connections/) ([Markdown](https://pllm.run/cli/reference/gateway/provider-connections.md)): Connect the trusted gateway to separately operated inference and preparation roles.
- [pllm Python API](https://pllm.run/sdk/reference/python/pllm/) ([Markdown](https://pllm.run/sdk/reference/python/pllm.md)): Public objects, practical usage, and research context for pllm.
- [pllm.client Python API](https://pllm.run/sdk/reference/python/pllm/client/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/client.md)): Public objects, practical usage, and research context for pllm.client.
- [pllm.compiler Python API](https://pllm.run/sdk/reference/python/pllm/compiler/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/compiler.md)): Public objects, practical usage, and research context for pllm.compiler.
- [pllm.components Python API](https://pllm.run/sdk/reference/python/pllm/components/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/components.md)): Public objects, practical usage, and research context for pllm.components.
- [pllm.config Python API](https://pllm.run/sdk/reference/python/pllm/config/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/config.md)): Public objects, practical usage, and research context for pllm.config.
- [pllm.correlation Python API](https://pllm.run/sdk/reference/python/pllm/correlation/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/correlation.md)): Public objects, practical usage, and research context for pllm.correlation.
- [pllm.deployment Python API](https://pllm.run/sdk/reference/python/pllm/deployment/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/deployment.md)): Public objects, practical usage, and research context for pllm.deployment.
- [pllm.evidence Python API](https://pllm.run/sdk/reference/python/pllm/evidence/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/evidence.md)): Public objects, practical usage, and research context for pllm.evidence.
- [pllm.kernels Python API](https://pllm.run/sdk/reference/python/pllm/kernels/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/kernels.md)): Public objects, practical usage, and research context for pllm.kernels.
- [pllm.metrics Python API](https://pllm.run/sdk/reference/python/pllm/metrics/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/metrics.md)): Public objects, practical usage, and research context for pllm.metrics.
- [pllm.models Python API](https://pllm.run/sdk/reference/python/pllm/models/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/models.md)): Public objects, practical usage, and research context for pllm.models.
- [pllm.native Python API](https://pllm.run/sdk/reference/python/pllm/native/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/native.md)): Public objects, practical usage, and research context for pllm.native.
- [pllm.nonlinear Python API](https://pllm.run/sdk/reference/python/pllm/nonlinear/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/nonlinear.md)): Public objects, practical usage, and research context for pllm.nonlinear.
- [pllm.official Python API](https://pllm.run/sdk/reference/python/pllm/official/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/official.md)): Public objects, practical usage, and research context for pllm.official.
- [pllm.passes Python API](https://pllm.run/sdk/reference/python/pllm/passes/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/passes.md)): Public objects, practical usage, and research context for pllm.passes.
- [pllm.pipeline Python API](https://pllm.run/sdk/reference/python/pllm/pipeline/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/pipeline.md)): Public objects, practical usage, and research context for pllm.pipeline.
- [pllm.plan Python API](https://pllm.run/sdk/reference/python/pllm/plan/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/plan.md)): Public objects, practical usage, and research context for pllm.plan.
- [pllm.preparation Python API](https://pllm.run/sdk/reference/python/pllm/preparation/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/preparation.md)): Public objects, practical usage, and research context for pllm.preparation.
- [pllm.profiles Python API](https://pllm.run/sdk/reference/python/pllm/profiles/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/profiles.md)): Public objects, practical usage, and research context for pllm.profiles.
- [pllm.protocols Python API](https://pllm.run/sdk/reference/python/pllm/protocols/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/protocols.md)): Public objects, practical usage, and research context for pllm.protocols.
- [pllm.protocols.masked_linear Python API](https://pllm.run/sdk/reference/python/pllm/protocols-masked-linear/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/protocols-masked-linear.md)): Public objects, practical usage, and research context for pllm.protocols.masked_linear.
- [pllm.providers Python API](https://pllm.run/sdk/reference/python/pllm/providers/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/providers.md)): Public objects, practical usage, and research context for pllm.providers.
- [pllm.research Python API](https://pllm.run/sdk/reference/python/pllm/research/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/research.md)): Public objects, practical usage, and research context for pllm.research.
- [pllm.roles Python API](https://pllm.run/sdk/reference/python/pllm/roles/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/roles.md)): Public objects, practical usage, and research context for pllm.roles.
- [pllm.runtime Python API](https://pllm.run/sdk/reference/python/pllm/runtime/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/runtime.md)): Public objects, practical usage, and research context for pllm.runtime.
- [pllm.schedulers Python API](https://pllm.run/sdk/reference/python/pllm/schedulers/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/schedulers.md)): Public objects, practical usage, and research context for pllm.schedulers.
- [pllm.search Python API](https://pllm.run/sdk/reference/python/pllm/search/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/search.md)): Public objects, practical usage, and research context for pllm.search.
- [pllm.server Python API](https://pllm.run/sdk/reference/python/pllm/server/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/server.md)): Public objects, practical usage, and research context for pllm.server.
- [pllm.sources Python API](https://pllm.run/sdk/reference/python/pllm/sources/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/sources.md)): Public objects, practical usage, and research context for pllm.sources.
- [pllm.state Python API](https://pllm.run/sdk/reference/python/pllm/state/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/state.md)): Public objects, practical usage, and research context for pllm.state.
- [pllm.verification Python API](https://pllm.run/sdk/reference/python/pllm/verification/) ([Markdown](https://pllm.run/sdk/reference/python/pllm/verification.md)): Public objects, practical usage, and research context for pllm.verification.
- [Component catalog](https://pllm.run/sdk/reference/components/) ([Markdown](https://pllm.run/sdk/reference/components.md)): Built-in component records with implementation, coverage, and evidence status.
- [Schema reference](https://pllm.run/sdk/reference/schemas/) ([Markdown](https://pllm.run/sdk/reference/schemas.md)): Versioned schemas for configurations, plans, evidence, assurance, and research records.
- [Current support](https://pllm.run/sdk/reference/status/) ([Markdown](https://pllm.run/sdk/reference/status.md)): What you can use today, what remains experimental, and what is not yet supported.
- [Source registry](https://pllm.run/research/sources/) ([Markdown](https://pllm.run/research/sources.md)): See how PLLM records original papers, source artifacts, citations, and licenses.
- [Method implementations](https://pllm.run/research/methods/) ([Markdown](https://pllm.run/research/methods.md)): Understand the difference between a published method and PLLM's independent implementation of it.
- [Compositions and recipes](https://pllm.run/research/compositions/) ([Markdown](https://pllm.run/research/compositions.md)): Learn how PLLM records research recipes and changes to a model plan.
- [Research evidence](https://pllm.run/research/evidence/) ([Markdown](https://pllm.run/research/evidence.md)): Understand which evidence supports fidelity, performance, privacy, and deployment claims.
- [Publications and claims](https://pllm.run/research/publications/) ([Markdown](https://pllm.run/research/publications.md)): Learn which evidence and review a PLLM research claim needs before publication.
- [Contribute a clean-room reimplementation](https://pllm.run/research/clean-room/) ([Markdown](https://pllm.run/research/clean-room.md)): Add an independent PLLM implementation with clear provenance, tests, evidence, and review.
- [PLLM research](https://pllm.run/research/) ([Markdown](https://pllm.run/research.md)): PLLM original research, technical paper, tracked papers, and implementation backlog.
- [PLLM: Composable Multi-Party Inference and Autonomous Research for Private LLMs](https://pllm.run/research/paper/) ([Markdown](https://pllm.run/research/paper.md)): Private multi-party LLM inference and evidence-driven research-component composition.
- [PLLM: Private LLM Inference That Keeps Plaintext with the User](https://pllm.run/research/whitepaper/) ([Markdown](https://pllm.run/research/whitepaper.md)): A concise guide to PLLM’s private multi-party runtime and autonomous research harness.
- [Implementation map](https://pllm.run/sdk/contribute/agents/implementation-map/) ([Markdown](https://pllm.run/sdk/contribute/agents/implementation-map.md)): Locate model planning, compiler contracts, native execution, Python APIs, CLI code, schemas, and documentation.
- [Agents](https://pllm.run/sdk/contribute/agents/) ([Markdown](https://pllm.run/sdk/contribute/agents.md)): Find the code and records used for implementation, research, and reproduction work.
- [Reproduction checklist](https://pllm.run/research/recipes/reproduction-checklist/) ([Markdown](https://pllm.run/research/recipes/reproduction-checklist.md)): A checklist for obtaining a source, reproducing a method, and producing comparable evidence.
- [Research map](https://pllm.run/research/recipes/agent-map/) ([Markdown](https://pllm.run/research/recipes/agent-map.md)): Trace a source through clean-room code, plans, evidence, and publication review.
- [Assurance](https://pllm.run/sdk/research/assurance/) ([Markdown](https://pllm.run/sdk/research/assurance.md)): Evaluate a specific privacy, integrity, numeric, implementation, or deployment claim.
- [Benchmarks](https://pllm.run/sdk/research/benchmarks/) ([Markdown](https://pllm.run/sdk/research/benchmarks.md)): Measure PLLM systems with reproducible records and compare only equivalent runs.
- [Benchmark and inspect runs](https://pllm.run/cli/benchmarking/) ([Markdown](https://pllm.run/cli/benchmarking.md)): Run the headless transport benchmark or open the diagnostic dashboard.
- [Command-line interface](https://pllm.run/cli/) ([Markdown](https://pllm.run/cli.md)): Inspect PLLM records and run local benchmarks from the terminal.
- [Inspect configuration and components](https://pllm.run/cli/inspect-and-research/) ([Markdown](https://pllm.run/cli/inspect-and-research.md)): Validate public experiment data and inspect built-in component records.
- [Run private inference](https://pllm.run/cli/private-inference/) ([Markdown](https://pllm.run/cli/private-inference.md)): Run the trusted gateway locally or connect it to independently operated PLLM roles.
- [Operate provider roles](https://pllm.run/cli/provider-roles/) ([Markdown](https://pllm.run/cli/provider-roles.md)): Start authenticated inference and preparation services without exposing application APIs.
- [Compiler](https://pllm.run/sdk/pipeline/compiler/) ([Markdown](https://pllm.run/sdk/pipeline/compiler.md)): Learn how PLLM turns a model plan and component choices into a checked execution plan.
- [Add a category](https://pllm.run/sdk/contribute/add-a-category/) ([Markdown](https://pllm.run/sdk/contribute/add-a-category.md)): Extend the semantic component graph when an existing category cannot express a real contract.
- [Component standard](https://pllm.run/sdk/contribute/component-standard/) ([Markdown](https://pllm.run/sdk/contribute/component-standard.md)): Required identity, lifecycle, representation, role, capability, evidence, and limitation fields.
- [Contribute](https://pllm.run/sdk/contribute/) ([Markdown](https://pllm.run/sdk/contribute.md)): Add a model, component, research method, or evidence record without creating a one-off runtime path.
- [Native plugin ABI](https://pllm.run/sdk/contribute/native-plugin-abi/) ([Markdown](https://pllm.run/sdk/contribute/native-plugin-abi.md)): C-compatible ownership, handles, status, and whole-region execution contracts for native providers.
- [Publish a provider](https://pllm.run/sdk/contribute/publish-a-provider/) ([Markdown](https://pllm.run/sdk/contribute/publish-a-provider.md)): Expose a component implementation without importing it during core discovery.
- [Upstreaming](https://pllm.run/sdk/contribute/upstreaming/) ([Markdown](https://pllm.run/sdk/contribute/upstreaming.md)): Move validated research work into maintained semantic and runtime components.
- [Conversions](https://pllm.run/sdk/build/conversions/) ([Markdown](https://pllm.run/sdk/build/conversions.md)): Learn how values move between numeric formats, layouts, and owners.
- [Deployment](https://pllm.run/sdk/operate/deployment/) ([Markdown](https://pllm.run/sdk/operate/deployment.md)): Learn how to assign PLLM roles to authenticated services without changing privacy assumptions.
- [Kernels](https://pllm.run/sdk/pipeline/kernels/) ([Markdown](https://pllm.run/sdk/pipeline/kernels.md)): Understand PLLM's native integer matrix kernels, memory costs, and CPU requirements.
- [Arithmetic and Boolean garbling](https://pllm.run/learn/arithmetic-and-boolean-garbling/) ([Markdown](https://pllm.run/learn/arithmetic-and-boolean-garbling.md)): Compare arithmetic and Boolean garbling without treating one primitive as a complete private runtime.
- [Learn](https://pllm.run/learn/) ([Markdown](https://pllm.run/learn.md)): Understand PLLM's private inference runtime, research harness, trust assumptions, and evidence.
- [Chat Completions API](https://pllm.run/learn/integrations/chat-completions/) ([Markdown](https://pllm.run/learn/integrations/chat-completions.md)): Connect text and local function-tool clients through PLLM's Chat Completions API adapter.
- [Codex](https://pllm.run/learn/integrations/codex/) ([Markdown](https://pllm.run/learn/integrations/codex.md)): Configure Codex CLI to use PLLM as a custom Responses API provider.
- [Integrations](https://pllm.run/learn/integrations/) ([Markdown](https://pllm.run/learn/integrations.md)): Connect OpenAI-compatible SDKs and coding agents to the trusted local PLLM gateway.
- [Local gateway](https://pllm.run/learn/integrations/local-gateway/) ([Markdown](https://pllm.run/learn/integrations/local-gateway.md)): Start the trusted loopback gateway locally or connect it to separate PLLM services.
- [OpenAI Agents SDK](https://pllm.run/learn/integrations/openai-agents/) ([Markdown](https://pllm.run/learn/integrations/openai-agents.md)): Run an OpenAI Agents SDK text agent through PLLM's Responses API gateway.
- [OpenAI Python SDK](https://pllm.run/learn/integrations/openai-python/) ([Markdown](https://pllm.run/learn/integrations/openai-python.md)): Point the official OpenAI Python SDK at the trusted PLLM gateway.
- [OpenCode](https://pllm.run/learn/integrations/opencode/) ([Markdown](https://pllm.run/learn/integrations/opencode.md)): Configure OpenCode to use PLLM through its Chat Completions API gateway.
- [Responses API](https://pllm.run/learn/integrations/responses-api/) ([Markdown](https://pllm.run/learn/integrations/responses-api.md)): Use PLLM's Responses API surface and understand its tested conformance scope.
- [Masked linear inference](https://pllm.run/learn/masked-linear-inference/) ([Markdown](https://pllm.run/learn/masked-linear-inference.md)): Follow seeded one-time masks through offline correction and online matrix evaluation.
- [Numeric semantics and model quality](https://pllm.run/learn/numeric-semantics-and-model-quality/) ([Markdown](https://pllm.run/learn/numeric-semantics-and-model-quality.md)): Understand how integer arithmetic affects correctness and model output quality.
- [Parties and offline work](https://pllm.run/learn/parties-and-offline-work/) ([Markdown](https://pllm.run/learn/parties-and-offline-work.md)): Learn what each party does before and during a private inference request.
- [Privacy and threat models](https://pllm.run/learn/privacy-and-threat-models/) ([Markdown](https://pllm.run/learn/privacy-and-threat-models.md)): Learn what PLLM protects, what each party can see, and which trust assumptions still apply.
- [Reading research and evidence](https://pllm.run/learn/reading-research-and-evidence/) ([Markdown](https://pllm.run/learn/reading-research-and-evidence.md)): Learn how PLLM connects papers, clean-room implementations, measurements, and publication claims.
- [Search](https://pllm.run/sdk/research/search/) ([Markdown](https://pllm.run/sdk/research/search.md)): Generate deterministic experiment candidates and rank only benchmark results with compatible evidence cohorts.
- [Metrics](https://pllm.run/research/records/metrics/) ([Markdown](https://pllm.run/research/records/metrics.md)): Choose clear performance, resource, quality, and cost measurements for a PLLM experiment.
- [Gemma 4](https://pllm.run/sdk/build/models/gemma4/) ([Markdown](https://pllm.run/sdk/build/models/gemma4.md)): Variant-parameterized text-decoder lowering for exact Gemma 4 E2B and E4B configurations.
- [Models](https://pllm.run/sdk/build/models/) ([Markdown](https://pllm.run/sdk/build/models.md)): See which model configurations PLLM can describe and what model support does not yet include.
- [Phi-4 mini](https://pllm.run/sdk/build/models/phi4-mini/) ([Markdown](https://pllm.run/sdk/build/models/phi4-mini.md)): Fused-projection and LongRoPE semantic lowering for Phi-4-mini-instruct.
- [Qwen2](https://pllm.run/sdk/build/models/qwen2/) ([Markdown](https://pllm.run/sdk/build/models/qwen2.md)): Dense Qwen2 semantic lowering and its present execution boundary.
- [Qwen3 and mini-coder](https://pllm.run/sdk/build/models/qwen3/) ([Markdown](https://pllm.run/sdk/build/models/qwen3.md)): Dense Qwen3 lowering with explicit head width and Q/K normalization.
- [Qwen3.5](https://pllm.run/sdk/build/models/qwen35/) ([Markdown](https://pllm.run/sdk/build/models/qwen35.md)): Hybrid full-attention and gated-delta semantic lowering for Qwen3.5-4B.
- [Numerics](https://pllm.run/sdk/build/numerics/) ([Markdown](https://pllm.run/sdk/build/numerics.md)): Understand PLLM's integer arithmetic, quantization, rounding, approximation, and overflow checks.
- [Python client](https://pllm.run/sdk/operate/client/) ([Markdown](https://pllm.run/sdk/operate/client.md)): Construct the native PLLM client, inspect its resources, and choose it instead of the gateway only when your process owns the private client boundary.
- [SiLU activation](https://pllm.run/sdk/build/operators/activation-silu/) ([Markdown](https://pllm.run/sdk/build/operators/activation-silu.md)): Compile and evaluate PLLM's bounded experimental Q7 SiLU component.
- [Operators](https://pllm.run/sdk/build/operators/) ([Markdown](https://pllm.run/sdk/build/operators.md)): Learn how PLLM describes model operations independently of model family and runtime.
- [Projection and rotary variants](https://pllm.run/sdk/build/operators/projection-rotary/) ([Markdown](https://pllm.run/sdk/build/operators/projection-rotary.md)): Represent fused projections, tensor slices, head widths, and rotary position rules.
- [Stateful recurrence](https://pllm.run/sdk/build/operators/stateful-recurrence/) ([Markdown](https://pllm.run/sdk/build/operators/stateful-recurrence.md)): Represent convolution and recurrent state in a hybrid decoder plan.
- [Pipeline](https://pllm.run/sdk/pipeline/) ([Markdown](https://pllm.run/sdk/pipeline.md)): Combine model, privacy protocol, preparation, runtime, and deployment choices in one configuration.
- [Preparation](https://pllm.run/sdk/pipeline/preparation/) ([Markdown](https://pllm.run/sdk/pipeline/preparation.md)): Prepare one-time masked material before a private inference request starts.
- [Arithmetic garbling](https://pllm.run/sdk/pipeline/protocols/garbling/arithmetic/) ([Markdown](https://pllm.run/sdk/pipeline/protocols/garbling/arithmetic.md)): Mixed-modulus labels and projection gates over bounded arithmetic values.
- [Half-gates](https://pllm.run/sdk/pipeline/protocols/garbling/half-gates/) ([Markdown](https://pllm.run/sdk/pipeline/protocols/garbling/half-gates.md)): Boolean AND-gate representation and the obligations surrounding free-XOR circuit execution.
- [Garbling](https://pllm.run/sdk/pipeline/protocols/garbling/) ([Markdown](https://pllm.run/sdk/pipeline/protocols/garbling.md)): Reference arithmetic and Boolean garbling components with explicit maturity and composition boundaries.
- [Lookup tables](https://pllm.run/sdk/pipeline/protocols/garbling/lookup-tables/) ([Markdown](https://pllm.run/sdk/pipeline/protocols/garbling/lookup-tables.md)): Bounded function tables with explicit domains, indexing, leakage, and material size.
- [Weighted path](https://pllm.run/sdk/pipeline/protocols/garbling/weighted-path/) ([Markdown](https://pllm.run/sdk/pipeline/protocols/garbling/weighted-path.md)): Decision-path representations for bounded piecewise or tree-structured secure evaluation.
- [Protocols](https://pllm.run/sdk/pipeline/protocols/) ([Markdown](https://pllm.run/sdk/pipeline/protocols.md)): See how PLLM defines parties, messages, privacy assumptions, and failure behavior.
- [Masked linear](https://pllm.run/sdk/pipeline/protocols/masked-linear/) ([Markdown](https://pllm.run/sdk/pipeline/protocols/masked-linear.md)): Prepared public-weight linear inference with seeded one-time masks and committed inventory.
- [Research workflows](https://pllm.run/research/recipes/) ([Markdown](https://pllm.run/research/recipes.md)): Build, reproduce, compare, and review private-inference research with explicit stop rules.
- [Native reference](https://pllm.run/sdk/reference/native/) ([Markdown](https://pllm.run/sdk/reference/native.md)): Rust crates, the Python binding, data ownership, threading, and ABI support.
- [Python reference](https://pllm.run/sdk/reference/python/) ([Markdown](https://pllm.run/sdk/reference/python.md)): The installed pllm namespace and its generated public API inventory.
- [Representations](https://pllm.run/sdk/build/representations/) ([Markdown](https://pllm.run/sdk/build/representations.md)): Understand how PLLM records a value's format, owner, bounds, and encoding.
- [Reimplementation backlog](https://pllm.run/research/backlog/) ([Markdown](https://pllm.run/research/backlog.md)): What PLLM will reproduce, implement, review, and measure next.
- [Experiments](https://pllm.run/research/recipes/experiments/) ([Markdown](https://pllm.run/research/recipes/experiments.md)): Design reproducible experiments with fixed plans, comparable runs, and recorded failures.
- [Research papers](https://pllm.run/research/papers/) ([Markdown](https://pllm.run/research/papers.md)): Prioritized paper reimplementation plan, source status, contribution summaries, and PLLM implementation boundaries.
- [R01: Dash: Accelerating Distributed Private Convolutional Neural Network Inference with Arithmetic Garbled Circuits](https://pllm.run/research/papers/r01-dash/) ([Markdown](https://pllm.run/research/papers/r01-dash.md)): Arithmetic garbling and LabelTensors provide the foundation for offline preparation and local encoded public-weight evaluation. PLLM status: reference primitive only.
- [R02: ReDASH: Fast and efficient Scaling in Arithmetic Garbled Circuits for Secure Outsourced Inference](https://pllm.run/research/papers/r02-redash/) ([Markdown](https://pllm.run/research/papers/r02-redash.md)): ReDASH generalizes residue-number-system scaling and uses ScaleQuant+ to reduce encoded-inference overhead. PLLM status: reference primitive only.
- [R03: Garbling Gadgets for Boolean and Arithmetic Circuits](https://pllm.run/research/papers/r03-garbling-gadgets/) ([Markdown](https://pllm.run/research/papers/r03-garbling-gadgets.md)): Arithmetic garbling provides free additions and specialized gadgets, with a stated symmetric-primitive security model. PLLM status: PLLM component adaptation.
- [R04: Two Halves Make a Whole: Reducing Data Transfer in Garbled Circuits using Half Gates](https://pllm.run/research/papers/r04-half-gates/) ([Markdown](https://pllm.run/research/papers/r04-half-gates.md)): Half-gates reduce AND-gate communication while remaining compatible with free XOR. PLLM status: tracked.
- [R05: Garbled Circuit Lookup Tables with Logarithmic Number of Ciphertexts](https://pllm.run/research/papers/r05-logrow/) ([Markdown](https://pllm.run/research/papers/r05-logrow.md)): The lookup construction reduces the number of security-parameter-sized ciphertexts; total communication retains a table-data term. PLLM status: tracked.
- [R06: Duty-Free Bits: Projectivizing Garbling Schemes](https://pllm.run/research/papers/r06-duty-free-bits/) ([Markdown](https://pllm.run/research/papers/r06-duty-free-bits.md)): This work studies projective conversion from Yao-style bit labels into affine encodings and related vector-OLE applications. PLLM status: tracked.
- [R07: Slalom: Fast, Verifiable and Private Execution of Neural Networks in Trusted Hardware](https://pllm.run/research/papers/r07-slalom/) ([Markdown](https://pllm.run/research/papers/r07-slalom.md)): Slalom outsources masked linear work and uses local verification within a trusted execution partition. PLLM status: tracked.
- [R08: Slalom at the Carnival: Privacy-preserving Inference with Masks from Public Knowledge](https://pllm.run/research/papers/r08-carnival/) ([Markdown](https://pllm.run/research/papers/r08-carnival.md)): Carnival uses subset-sum-based masks to accelerate preprocessing for linear outsourcing. PLLM status: tracked.
- [R09: HyCC: Compilation of Hybrid Protocols for Practical Secure Computation](https://pllm.run/research/papers/r09-hycc/) ([Markdown](https://pllm.run/research/papers/r09-hycc.md)): HyCC compiles high-level programs into hybrid secure-computation protocols. PLLM status: tracked.
- [R10: Oblivious decision program evaluation](https://pllm.run/research/papers/r10-oblivious-decision-programs/) ([Markdown](https://pllm.run/research/papers/r10-oblivious-decision-programs.md)): Oblivious decision-program protocols evaluate protected decision structures using OT-based reductions. PLLM status: tracked.
- [R11: Factored Edge-Valued Binary Decision Diagrams](https://pllm.run/research/papers/r11-fevbdd/) ([Markdown](https://pllm.run/research/papers/r11-fevbdd.md)): Factored edge-valued diagrams share subfunctions through additive and multiplicative edge normalization. PLLM status: tracked.
- [R12: Piranha: A GPU Platform for Secure Computation](https://pllm.run/research/papers/r12-piranha/) ([Markdown](https://pllm.run/research/papers/r12-piranha.md)): Piranha separates device, protocol and application layers for GPU-accelerated secret-sharing computation. PLLM status: tracked.
- [R13: SIGMA: Secure GPT Inference with Function Secret Sharing](https://pllm.run/research/papers/r13-sigma/) ([Markdown](https://pllm.run/research/papers/r13-sigma.md)): SIGMA provides FSS-based secure transformer operations and GPU-backed inference with preprocessing. PLLM status: tracked.
- [R14: FuseFSS: Efficient Secure LLM Inference with Function Secret Sharing](https://pllm.run/research/papers/r14-fusefss/) ([Markdown](https://pllm.run/research/papers/r14-fusefss.md)): FuseFSS targets fusion and efficiency of secure nonlinear/helper operations in FSS-based LLM inference. PLLM status: tracked.
- [R15: Efficient Pseudorandom Correlation Generators over Z/p^k Z](https://pllm.run/research/papers/r15-ring-pcg/) ([Markdown](https://pllm.run/research/papers/r15-ring-pcg.md)): Galois-ring constructions extend pseudorandom correlation generation to integer rings, including multiplication-related correlations. PLLM status: tracked.
- [R16: Practical Secure Delegated Linear Algebra with Trapdoored Matrices](https://pllm.run/research/papers/r16-trapdoored-matrices/) ([Markdown](https://pllm.run/research/papers/r16-trapdoored-matrices.md)): This work uses trapdoored pseudorandom matrices for efficient delegated linear algebra under an LPN-based approach. PLLM status: tracked.
- [R17: Curl: Private LLMs through Wavelet-Encoded Look-Up Tables](https://pllm.run/research/papers/r17-curl/) ([Markdown](https://pllm.run/research/papers/r17-curl.md)): Curl uses wavelet-encoded lookup tables for private nonlinear operations and discusses probabilistic truncation security. PLLM status: tracked.
- [R18: Compact: Approximating Complex Activation Functions for Secure Computation](https://pllm.run/research/papers/r18-compact/) ([Markdown](https://pllm.run/research/papers/r18-compact.md)): Compact builds input-density-aware piecewise polynomial approximations for complex activations. PLLM status: tracked.
- [R19: Private and Verifiable Outsourcing of Open-Weight LLM Inference](https://pllm.run/research/papers/r19-open-weight-verifiable/) ([Markdown](https://pllm.run/research/papers/r19-open-weight-verifiable.md)): The abstract describes private and verifiable open-weight inference with two malicious, non-colluding servers and model-owner-published information. PLLM status: tracked.
- [R20: Oblivious Ciphertext Compression via Linear Codes](https://pllm.run/research/papers/r20-oblivious-compression/) ([Markdown](https://pllm.run/research/papers/r20-oblivious-compression.md)): This work develops oblivious ciphertext compression using linear codes and syndrome decoding. PLLM status: tracked.
- [R21: NEXUS: Secure Transformer Inference Made Non-interactive](https://pllm.run/research/papers/r21-nexus/) ([Markdown](https://pllm.run/research/papers/r21-nexus.md)): NEXUS combines approximate HE with lightweight garbled circuits for non-interactive two-party transformer forward inference. PLLM status: tracked.
- [R22: BumbleBee: Secure Two-party Inference Framework for Large Transformers](https://pllm.run/research/papers/r22-openbumblebee/) ([Markdown](https://pllm.run/research/papers/r22-openbumblebee.md)): BumbleBee implements secure two-party transformer inference with specialized nonlinear protocols and an Apache-2.0 artifact. PLLM status: tracked.
- [R23: MPCache: MPC-Friendly KV Cache Eviction for Efficient Private Large Language Model Inference](https://pllm.run/research/papers/r23-mpcache/) ([Markdown](https://pllm.run/research/papers/r23-mpcache.md)): MPCache evaluates MPC-friendly static and dynamic KV-cache eviction in a three-party SecretFlow/SPU setting; its artifact focuses state selection and operator costs. PLLM status: PLLM reimplementation.
- [R24: Maverick: Private and Verifiable LLM Inference Made Practical via Matrix-Vector Multiplication Delegation](https://pllm.run/research/papers/r24-maverick/) ([Markdown](https://pllm.run/research/papers/r24-maverick.md)): Maverick delegates matrix-vector multiplication with transparent preprocessing, information-theoretic batch verification, and LPN-based pseudorandom input masking while adding virtually no server-side verification overhead. PLLM status: planned.
- [Evidence and records](https://pllm.run/research/records/) ([Markdown](https://pllm.run/research/records.md)): Keep sources, methods, implementations, plans, evidence, assurance, and claims independently reviewable.
- [Reproductions](https://pllm.run/research/recipes/reproductions/) ([Markdown](https://pllm.run/research/recipes/reproductions.md)): Reimplement published methods independently in PLLM's Python and Rust stack.
- [Runtime](https://pllm.run/sdk/pipeline/runtime/) ([Markdown](https://pllm.run/sdk/pipeline/runtime.md)): Understand which live state a PLLM runtime owns and how it differs from a reproducible plan.
- [Python SDK](https://pllm.run/sdk/) ([Markdown](https://pllm.run/sdk.md)): Plan models, compose research methods, and use supported PLLM runtimes from Python.
- [Search](https://pllm.run/sdk/build/search/) ([Markdown](https://pllm.run/sdk/build/search.md)): Compare compatible system designs against explicit resource, quality, and assurance limits.
- [Run the local development benchmark](https://pllm.run/learn/start/first-local-benchmark/) ([Markdown](https://pllm.run/learn/start/first-local-benchmark.md)): Run the real PLLM roles on one machine and understand what the result does and does not measure.