LMRuntime.com / Current state

Project Status

A plain-language view of the published package family, generated documentation, version policy, and evidence boundaries that remain application-specific.

The public site documents 17 UAIX.LmRuntime packages, 77 source-grounded examples, and a generated API reference. NuGet remains authoritative for current package release metadata.

Works Today

The package family is publicly discoverable on NuGet, with dedicated guides and repeated direct NuGet links on this site.
LocalEndpoint provides the verified local GGUF managed CPU application path.
Acceleration defines backend registration, probing, selection policies, device descriptors, diagnostics, and visible CPU fallback state.
Backends.CpuManaged reports available without a native dependency.
GGUF, tokenization, sampling, tensors, CPU kernels, and LLaMA execution have focused package surfaces and examples.

Backend state

Package/API state

CUDA, DirectML, Vulkan, ROCm, and Metal registration packages are public.
Each declares identity, compatibility metadata, and diagnostic behavior.
CUDA native-asset identities are separate packages.

Execution state not inferred

A GPU registration package does not by itself provide GPU inference.
A declared runtime identifier is not a successful device probe.
The supplied source manifests do not establish an embedded CUDA inference binary in the native-asset slots.

Documentation state

Each package guide includes role and boundary text, version-free install commands, direct dependencies, key types, deep C# examples, FAQs, a generated member reference, and links to the package's NuGet page. The API dataset represents 376 public types and 2,044 documented members.

Version policy

Public documentation deliberately omits UAIX.LmRuntime package version numbers. Applications should choose and pin resolved versions through normal dependency-management and lock-file processes, then verify upgrades in CI. NuGet is the current source for target frameworks, dependency versions, package hashes, downloads, and release dates.

Experimental roadmap

Research proposal

Evidence-Gated Adaptive Runtime Control

A host-layer experiment would combine explicit budgets, backend-selection evidence, bounded recovery, session restoration, no-op decisions, and reproducible records around the existing package family. It does not change package source or claim autonomous runtime behavior.

Read the roadmap Browse packages on NuGet ↗

Not Claimed

Universal GGUF supportContainer validity, tokenizer compatibility, model architecture, tensor storage, and kernel coverage are separate checks.

GPU inference through registration aloneExecution requires a successful local probe and a proven adapter, not only a package reference.

Benchmark leadershipNo speed, latency, throughput, memory, or energy claim is made without a reproducible environment and retained results.

Cloud or server platformThe family does not imply a hosted service, model download service, provider client, or network fallback.

Production or safety certificationEach consuming application owns deployment review, controls, supported-model declarations, operational testing, and support terms.