Drop-in local provider for Semantic Kernel and Microsoft.Extensions.AI
Official bridge packages let LM-Kit.NET act as the local model provider for any existing Semantic Kernel or Microsoft.Extensions.AI application.
Free download
Get LM-Kit.NET up and running in minutes.
Install via NuGet, pick a model, and build your first AI-powered .NET application. Full feature access, no time limits, no registration required, zero cloud dependency.
From zero to AI in three steps.
Create a new project, install the NuGet package, and follow the getting-started guide to build your first AI application.
01 · Project
Create a .NET project
Start from a console app or add LM-Kit.NET to any existing .NET solution.
dotnet new console -n MyAIApp
02 · Install
Install the NuGet package
Single dependency, in-process. Includes optimised CPU inference out of the box.
dotnet add package LM-Kit.NET
03 · Build
Build your first AI app
Initialise the runtime, load a model, generate a response. The getting-started guide walks every step.
Your first AI agentOne command. NuGet does the rest.
LM-Kit.NET ships as a single NuGet package with optimised CPU inference (AVX/AVX2) included. Add an optional backend package for GPU acceleration when you need it.
.NET CLI
dotnet add package LM-Kit.NET
From any terminal, in any project directory. Equivalent in the Package Manager Console: Install-Package LM-Kit.NET
CUDA 12 · Win
NVIDIA GPU on Windows
dotnet add package LM-Kit.NET.Backend.Cuda12.Windows
CUDA 12 · Linux
NVIDIA GPU on Linux
dotnet add package LM-Kit.NET.Backend.Cuda12.Linux
Vulkan & Metal
Cross-vendor GPU & Apple Silicon
Both backends are bundled in the main package. No extra install required.
Choose the right model for your use case.
The catalog is continuously updated with releases from Google, Meta, Alibaba, Mistral, Microsoft and IBM. Models auto-download on first use.
Chat & code
General conversation and coding
Gemma, Qwen, Llama, Phi, Mistral, Granite, Falcon, SmolLM. Sub-1B to 30B+.
Reasoning & agents
Multi-step reasoning, tool use
Qwen 3.8, GPT-OSS, GLM 4.7 Flash, Nemotron 3 Nano, Ministral 3.
Vision & multimodal
VLMs and OCR
Qwen-VL, MiniCPM, Ministral, Pixtral, Devstral, plus dedicated OCR models.
OCR
Document text extraction
PaddleOCR, LightOnOCR. Structured text from images and scanned documents.
Embeddings
Vector search and reranking
Qwen Embed, EmbeddingGemma, Harrier OSS, BGE, Nomic. Text and image embeddings, plus rerankers.
Speech
High-accuracy audio transcription
A growing local speech-to-text catalog. Multiple model sizes for different speed and accuracy trade-offs.
What you can build.
A single SDK covering the full surface of AI features. Everything runs 100% on-device, zero cloud dependency. Explore all use cases →
Agents
AI agent orchestration
Multi-agent workflows with persistent memory, reasoning, function calling, MCP integration. Pipeline, parallel, supervisor, router patterns.
AgentsDocuments
Document intelligence
Chat with PDFs, extract structured data, convert documents, perform OCR on images.
Document intelligenceRAG
RAG & knowledge
Built-in vector database, semantic search, intelligent splitting, citation tracking.
RAG & knowledgeVision & speech
Multimodal understanding
Image understanding with vision-language models. High-accuracy on-device speech-to-text.
Speech to textText Analysis
Classification, sentiment, NER, PII
Sentiment, classification, translation, summarisation, named-entity recognition, PII extraction.
Text AnalysisText generation
Content, rewriting, structured output
Generation, rewriting, grammar correction, JSON-schema-constrained structured output.
Text generationChatbots
Conversational assistants
Multi-turn agents with persistent memory, streaming responses, agent skills, human escalation.
ChatbotsMCP & tools
Tool integration
Connect agents to external services via Model Context Protocol. 56 built-in tools for file I/O, HTTP, web search.
MCP & toolsWorks everywhere you build .NET.
Targets .NET Standard 2.0 for maximum compatibility. Develop and deploy on the platform of your choice.
- Operating systems
- Windows x64 · macOS (Intel & Apple Silicon) · Linux x64 / ARM64
- GPU acceleration
- CUDA 12 / 13 · Vulkan · Metal · CPU AVX/AVX2 fallback
- .NET frameworks
- .NET Standard 2.0 · .NET 8 / 9 / 10
Frequently asked questions.
Do I need a GPU to use LM-Kit.NET?
No. LM-Kit.NET includes optimised CPU inference (AVX/AVX2) out of the box, fine for smaller models and lower-throughput scenarios. For production workloads or larger models, GPU acceleration via CUDA, Vulkan or Metal significantly improves performance. See the GPU setup guide.
Is it really free?
Yes, with no time limit, no feature lock, no registration and no license key. Evaluation and development are free at any company size. Commercial use and redistribution are free for companies under $1M USD annual gross revenue, with 10 or fewer employees, that have raised no more than $3M USD from outside investors. Above that, see the pricing page.
What models are supported?
A continuously-updated catalog of open-weight models from Google, Meta, Alibaba, Mistral, Microsoft and IBM. Chat, reasoning, vision, OCR, embedding and speech models. Models auto-download on first use. Browse the full catalog.
Does it work offline or in air-gapped environments?
Yes. Once a model is downloaded, all inference runs 100% on-device with zero cloud dependency. Pre-download models and deploy in fully air-gapped environments. Ideal for regulated industries, healthcare, defence, and any scenario requiring complete data sovereignty.
How does local inference compare to cloud APIs?
Local inference eliminates per-token API costs, avoids rate limits, and keeps all data on your infrastructure. See the cost and performance comparison.
Can I integrate with Microsoft Semantic Kernel or Extensions.AI?
Yes. Official bridge packages exist for both Semantic Kernel and Microsoft.Extensions.AI.
Nothing to activate.
Install the package and build. No key, no signup, no time limit, no feature locks. A Professional license applies only once your company passes the free thresholds.