LM-Kit.NET

Each of these pillars is designed to function seamlessly on the core runtime. The LM-Kit.NET NuGet package includes the entire inference system, featuring open-weight LLMs, vision-language models, embedding models, on-device speech-to-text, OCR, and classifiers. These components are accelerated across various hardware, including CPU, AVX2, CUDA 12/13, Vulkan, and Metal, ensuring efficient local execution. This approach guarantees predictable latency, full data sovereignty, and technology independence, as there are no cloud calls involved.

The AI Agents pillar supports advanced orchestration patterns like ReAct planning, supervisors, and parallel/pipeline orchestrators, along with persistent memory and custom tools. Document Intelligence offers robust PDF and image parsing, on-device OCR achieving state-of-the-art benchmark scores, and structured field extraction using grammar-constrained generation. The Vision & Multimodal capabilities include image understanding, classification, labeling, multimodal chat, and VLM-OCR, offering the same conversational interface as LLMs.

For RAG & Knowledge, LM-Kit.NET includes a built-in vector store with connectors for Qdrant and pgvector, supporting hybrid retrieval and source citations. Text Analysis provides built-in classifiers and extractors for sentiment, named entity recognition (NER), PII detection, and language detection, emitting typed C# objects. The Speech & Audio pillar features a growing local speech-to-text stack with hallucination suppression, Voice Activity Detection, real-time translation, and streaming output for over 100 languages.

Text Generation encompasses single-turn, multi-turn, and stateless conversation primitives, including translation, rewriting, summarization, prompt templates, and grammar-constrained outputs. The foundation of LM-Kit.NET is its adaptive inference engine, which features constrained output via dynamic grammar for guaranteed parsing of JSON, schemas, and tool-call shapes, adaptive guidance for reduced hallucinations, and model-agnostic architecture for easy integration of new open-weight models.

LM-Kit.NET offers significant advantages over cloud APIs, including no per-token billing, no data leaving the network, predictable latency, and leveraging existing compute infrastructure. It also contrasts with Python services by avoiding sidecars and HTTP shims, integrating directly into the .NET process. Compared to other .NET local stacks, LM-Kit.NET provides the full runtime, not just inference, encompassing agents, RAG, OCR, speech, vision, and classifiers.

LM-Kit.NET is compatible with .NET Standard 2.0 and .NET 8/9/10, running on Windows, Linux (x64 & ARM64), and macOS. It offers a free community license for non-commercial applications and evaluation, with a professional, custom-priced option for commercial redistribution. Bridges are available for Microsoft.Extensions.AI and Semantic Kernel, allowing seamless integration with existing .NET AI ecosystems.