LM-Kit.NET · Private Document Intelligence

The Private Document Intelligence SDK for .NET.

Extract structured data, redact PII, and answer questions over company documents with citations, embedded directly in your .NET application. Zero cloud calls.

Quickstart Free Download

Prefer a server you deploy over a runtime you embed? See LM-Kit One →

One NuGet .NET Standard 2.0 · .NET 8 / 9 / 10 Windows · Linux · macOS

From documents to data you can verify.

The document stack reads PDF, Office, email, HTML and scans. Extraction returns schema-valid JSON with a confidence figure per field; redaction permanently removes what a reviewer approves.

Extract

Typed fields with per-field confidence

Schema-driven extraction with grammar-constrained generation. Invoices, contracts, forms, IDs. Every field carries a confidence score and a human-review flag.

Structured extraction

Convert

Any document to LLM-ready Markdown

PDF, DOCX, PPTX, XLSX, EML, HTML and scans become clean Markdown, with per-page routing between text extraction and VLM OCR.

Document to Markdown

OCR

Native and VLM-driven OCR

CPU-efficient native engine plus VLM OCR (PaddleOCR-VL 1.6, LightOnOCR-2). Fast on a single core, no per-page cost, no OCR licence to budget for.

OCR

Split

Detect boundaries in 1000-page batches

Intelligent document splitting finds logical boundaries in long PDFs. Cut a batch into per-document files automatically.

Document splitting

Redact

PII redaction a reviewer approves

Detect sensitive content, review every finding, and permanently remove approved regions. True text removal, not a black rectangle drawn over it.

Smart Redaction

Operate

Split, merge, archive, protect

Full PDF toolkit: split, merge, search, render, unlock, and convert to archival-grade PDF/A. The file operations around the intelligence.

PDF toolkit

From documents to knowledge that cites its source.

Ask questions over manuals, policies, contracts and wikis. Retrieval, reranking and grounded generation return the document, the page and the passage behind every answer.

RAG & knowledge →

  • Multi-document chat. PdfChat loads several documents into one conversation with automatic context management.
  • Page-level citations. DocumentRag pairs OCR with VLM understanding and returns source references on every answer.
  • Hybrid retrieval. Vector, BM25 and hybrid search with MMR diversity and cross-encoder reranking.
  • Pluggable storage. In-memory, built-in file DB, Qdrant or PostgreSQL (pgvector) behind one IVectorStore contract.
  • Text and image embeddings. Batched, async-first, with cross-modal similarity out of the box.

Fits the AI stack you already run.

Bridges for the .NET AI abstractions, connectors for production vector databases, a provider-agnostic embedding layer, and MCP. Keep your orchestration and storage; move inference local where it counts.

Bridge

Microsoft.Extensions.AI

Stream tokens, call functions, embed text. Every IChatClient, IEmbeddingGenerator, and middleware-aware abstraction you wrote against the official package keeps working with LM-Kit as the local backend.

Extensions.AI bridge

Bridge

Semantic Kernel

Use LM-Kit.NET as a Semantic Kernel connector. Plug local chat completion, embeddings, and function-calling into existing SK plans, planners, and skills.

Semantic Kernel bridge

Embeddings

Amazon Bedrock

Bring a managed embedding provider to your RAG pipeline. Amazon Titan and Cohere models on Bedrock plug in through the same IEmbedder contract as local models.

Amazon Bedrock bridge

Connectors

Vector databases

Persist RAG embeddings in production stores. Official connectors for Qdrant and PostgreSQL/pgvector implement the same IVectorStore as the built-in memory path.

Vector database connectors

Protocol

Model Context Protocol

MCP clients ship in the agent runtime. Connect to any MCP server; built-in tools and external MCP tools coexist in a unified registry with the same permission policy.

MCP integration

One package. The whole AI stack.

Document intelligence needs more than a model. The same NuGet that runs extraction also ships OCR, speech, vision, agents, RAG, classifiers, and embeddings. No Python sidecar, no Docker.

Models

100+

Pre-configured cards plus any GGUF from Hugging Face.

Pillars

7

Agents, Docs, Vision, RAG, Text, Speech, Generation.

Built-in tools

8 categories

Atomic, security-first tools. Constantly growing catalog.

Backends

5

CPU, AVX2, CUDA 12/13, Vulkan, Metal. Same code path.

Cloud calls

0

Every model runs on your hardware. No data leaves the box.

External services

0

In-process SDK. No Python runtime, no Docker, no daemons.

Vector backends

5

In-memory, built-in file DB, Qdrant, PostgreSQL (pgvector), bring-your-own.

Speech languages

100+

Whisper-family STT with VAD and hallucination suppression.

Seven pillars, one foundation.

The seven pillars of LM-Kit.NET, plus the local runtime they share. Highlighted card is where you are now.

01 · AI Agents

Orchestration patterns

ReAct planning, supervisors, parallel and pipeline orchestrators, persistent memory, MCP clients, custom tools.

AI Agents

02 · Document Intelligence

Parse PDFs, images, EML

PDF text and table extraction, on-device OCR with layout and table recovery, structured field extraction with grammar-constrained generation.

Document Intelligence

03 · Vision & Multimodal

VLMs, image classification, chat with image

Image understanding, classification, labeling, multimodal chat, image embeddings, VLM-OCR, background removal. Same conversation surface as LLMs.

Vision & Multimodal

04 · RAG & Knowledge

Vector search and retrieval

Built-in vector store, Qdrant and pgvector connectors, embeddings, hybrid retrieval, document chunking, source citations.

RAG & Knowledge

05 · Text Analysis

Classification, NER, PII, sentiment

Built-in classifiers and an extractor that emits typed C# objects via grammar-constrained sampling. Sentiment, keywords, language detection.

Text Analysis

06 · Speech & Audio

Audio transcription, STT

A growing local speech-to-text stack: hallucination suppression, Voice Activity Detection, real-time translation, streaming output, 100+ languages.

Speech & Audio

07 · Text Generation

Conversations, rewriting, summaries

Single-turn, multi-turn, and stateless conversation primitives. Translate, correct, rewrite, summarise. Prompt templates, streaming, grammar-constrained outputs.

Text Generation

The foundation

Every capability above runs on this runtime.

Foundation

Local Inference

The runtime all seven pillars sit on. The LM-Kit.NET NuGet ships the complete inference system: open-weight LLMs, vision-language models, embeddings, on-device speech-to-text, OCR and classifiers, accelerated on CPU, AVX2, CUDA 12/13, Vulkan or Metal. One package, zero cloud calls, predictable latency, full data and technology sovereignty.

Explore the foundation

Dynamic Sampling, the symbolic layer.

Underneath every LM-Kit call sits an adaptive inference engine that steers each token in real time using structural awareness, contextual signals, and grammar-aligned validation. It is the reason a 4B local model can match fine-tuned cloud behaviour on extraction, classification, function calling, and structured generation. Always on, model-agnostic, no retraining required.

Pillar A

Constrained output

Dynamic grammar guarantees JSON, schemas, and tool-call shapes always parse. A hybrid path keeps it measurably faster than classical grammar sampling.

Pillar B

Adaptive guidance

Per-token contextual perplexity, semantic memory for codes and identifiers, structural rejection of malformed runs. Hallucinations drop, recoveries happen in place.

Pillar C

Model-agnostic

No architecture coupling, no fine-tuning, no per-model adapter. Drop in a new open-weight release and the layer keeps working from day one.

Open the Dynamic Sampling deep dive →

Eight categories of atomic, security-first tools.

LM-Kit ships a growing catalog of agent tools across eight categories. Each tool performs exactly one operation, exposes rich metadata, and integrates with the permission policy system for policy-driven access control. One tool, one feature. Compose freely.

01 · Data

Parse and transform

JSON, XML, CSV, YAML, HTML, Markdown, databases, spreadsheets, QR codes. Predictable, typed I/O.

02 · Document

PDF, OCR, format conversion

PDF manipulation, image preprocessing, OCR, format conversion between Markdown, EML, HTML, DOCX.

03 · Text

String operations

Diff, regex, templating, encoding, slugification, fuzzy matching, phonetics. The stuff prompts cannot do.

04 · Numeric

Compute primitives

Calculator, unit conversion, statistics, financial math, random, expression evaluation.

05 · Security

Hashing, JWT, validation

Hashing, encryption, JWT, validation, password generation, checksums. Audit-friendly defaults.

06 · Utility

Date, URL, locale, MIME

Date and time, cron, URLs, colors, locales, MIME types, paths, scheduling, time zones.

07 · IO

Filesystem, processes, files

File system, process execution, compression, clipboard, environment, file watching.

08 · Net

HTTP, web search, RSS

HTTP verbs, FTP, web search (DuckDuckGo, Brave, Tavily, Serper, SearXNG), SMTP, RSS feeds, diagnostics.

Permission policies

Every tool implements IToolMetadata with explicit risk level, side effect kind (LocalRead / LocalWrite / NetworkRead / NetworkWrite / Irreversible), default approval mode, and read-only flag. Pair with ToolPermissionPolicy for centralized allow/deny rules, wildcard patterns, and approval gates. Production-safe out of the box.

Open-weight models, curated and ready.

A constantly growing catalog of open-weight models covering text generation, vision, embeddings, OCR, and speech. Load any of them by ID, or point at any GGUF on Hugging Face.

Text LLMs

From 0.6B to 30B-class MoE

Qwen 3.8, Gemma 4, GLM 4.7 Flash, GPT-OSS 20B, Granite 4 and Ministral 3, from 0.8B to 31B. Chat, reasoning with adjustable effort, tool use, multilingual.

Vision

VLMs for chat and OCR

Qwen3-VL (2B to 30B), Muse Glimmer 30B, GLM-V 4.6 Flash. Dedicated OCR via PaddleOCR-VL 1.6, LightOnOCR-2 and Infinity Parser 2. Drop an Attachment into any conversation.

Embeddings

Text and image vectors

Qwen3-Embedding 0.6B / 4B / 8B and Harrier OSS 0.6B for text, Nomic-Embed-Vision for cross-modal retrieval. Multilingual, with reranking.

Speech

Whisper family STT

Tiny through Large V3 and Large Turbo V3. 100+ languages, real-time translation to English, Voice Activity Detection.

Task models

Purpose-trained classifiers

Sentiment-analysis 2.0 and lmkit-tasks variants for fast on-device classification, NER, and PII work.

Bring-your-own

Any GGUF, any URI

Point new LM(uri) at any GGUF on Hugging Face or your own storage. The catalog is curation, not constraint.

Browse the catalog Pick a model

From knowledge to governed action.

A strongly-typed agent runtime with planning strategies, tool registries behind permission policies, persistent memory, MCP clients, multi-agent orchestration, and production-grade observability. Compose freely, ship confidently.

Agent runtime

Strongly-typed agents

First-class Agent with builders, identity, tools, planning strategy, retry policies, streaming, and event-level observability.

AI Agents

Planning

ReAct, single-call, plan-then-execute

Pick the planning strategy per agent. Bounded steps, recoverable errors, deterministic by configuration.

Agent reasoning

Orchestration

Pipeline, parallel, supervisor

Sequential pipelines, parallel fan-out, supervisor delegation. Compose agents into deterministic graphs.

Multi-agent workflows

Graph

Graph orchestration

Sequential, Parallel, and Conditional nodes. Thread-safe context, channel-based streaming, compose any workflow shape.

Graph orchestration

Memory

Persistent agent memory

RAG-backed long-term memory that survives sessions. Recall, summarisation, semantic retrieval out of the box.

Agent memory

Skills

Reusable SKILL.md files

Drop SKILL.md files into a folder. The registry picks them up. Versionable, testable, swappable.

Agent skills

MCP

Model Context Protocol

Connect to any MCP-compatible server. The built-in tool catalog and external MCP tools coexist in the same registry.

MCP integration

Guardrails

Permissions and approval gates

Centralized policy: allow/deny patterns, max risk level, approval requirements per tool. Audit-friendly, production-safe.

Permissions

Observability

Trace every step

Events for plan, tool call, model decision, retry, error. Integrate with your existing telemetry pipeline.

Observability

Vision, speech, and content intelligence, same runtime.

LM-Kit.NET is not just an LLM library. The same NuGet ships production vision-language models, speech-to-text with Voice Activity Detection, sentiment and emotion classifiers, NER and PII extractors, multilingual translation, and grammar-aware text correction.

Vision

Chat with images

Drop an Attachment into MultiTurnConversation. Multiple images per turn, streaming tokens, tool calls. Image understanding, classification, labeling, background removal.

Vision & multimodal

Speech

Transcription, translation, dictation

Whisper-family STT. 100+ languages, real-time translation, Voice Activity Detection, hallucination suppression, voice-command dictation formatting.

Speech & audio

Sentiment

Sentiment & emotion

Multilingual polarity and emotion classification with neutral support. LoRA-fine-tunable on your domain.

Sentiment

NER

Custom entity recognition

Built-in entity types plus custom EntityDefinition. Character offsets on every mention for precise downstream use.

NER

PII

PII detection & redaction

Emails, phones, IDs, addresses, custom domain labels. Batch processing for high-volume document scanning.

PII extraction

Generation

Conversations, summaries, rewriting

Single-turn, multi-turn, and stateless primitives. Translate, correct, rewrite, summarise. Prompt templates, streaming, grammar-constrained outputs.

Text generation

The boring parts that ship.

Beyond the headline features, LM-Kit.NET ships the production controls a team needs once a prototype meets real workloads: memory hibernation, encrypted model loading, multi-GPU split, LoRA, fine-tuning, quantization, and sampling levers.

Hibernation

Context hibernation

Serialize an entire conversation context (KV-cache and all session state) to disk. Free RAM/VRAM on demand. Rehydrate transparently on the next call.

Context hibernation

Encryption

Encrypted model loading

Stream-decrypted GGUF. No plaintext model files on disk at any point. Ship the model, keep the secret.

Encrypted models

Multi-GPU

Multi-GPU and tensor overrides

Split big models across multiple GPUs. Per-tensor placement overrides for fine-grained control on heterogeneous hardware.

Multi-GPU

Sampling

Temperature, top-k, top-p, grammar

Per-turn sampling levers composed on top of Dynamic Sampling. Switch from creative to deterministic without rebuilding the conversation.

Sampling controls

Fine-tuning

In-process LLM fine-tuning

Bring data, build a LoRA, or fine-tune the full model. The training loop runs in the same NuGet that runs inference.

LLM fine-tuning

LoRA

LoRA adapter loading

Compose multiple LoRA adapters at load time. Swap personas, domains, or tasks without reloading the base model.

LoRA integration

Quantization

In-process quantization

Quantize models to fit the hardware budget without leaving the SDK. From Q8 down to Q2_K, with knobs for precision-critical layers.

Model quantization

Backends

CPU, AVX2, CUDA, Vulkan, Metal

Same C# call dispatches to the fastest available backend. Deploy once, run on anything from a Raspberry Pi to an 8-GPU server.

Hardware backends

Cross-platform by default, not by accident.

Same NuGet, same API surface, every supported target. Targets .NET Standard 2.0 so it slots into existing .NET Framework 4.6.2+ codebases too.

Runtime
.NET Standard 2.0 · .NET 8 · 9 · 10
Operating systems
Windows 10+ · Linux x64 & ARM64 · macOS Universal
GPU acceleration
CUDA 12 · CUDA 13 · Vulkan · Metal
CPU acceleration
SSE 4.1 / 4.2 · AVX · AVX2
Models
Qwen 3.8, Gemma 4, GLM 4.7, Muse Glimmer, GPT OSS, Granite 4, plus VLMs, OCR, embeddings and speech
Storage
In-memory · built-in file DB · Qdrant · pgvector · bring-your-own
Languages
C# · F# · VB.NET · any .NET-compatible language
Integrations
Microsoft.Extensions.AI · Semantic Kernel · Amazon Bedrock · Qdrant · pgvector · MCP

Pick the right tradeoff for your stack.

Three honest comparisons with the alternatives a .NET team actually weighs. No straw men, no invented numbers.

Compare

vs. wrapping cloud APIs

No per-token bill. No data leaving your network. Latency you can predict. Inference cost equals the cost of compute you already own. Works offline by design.

Local vs Cloud, in depth

Compare

vs. a Python service

No FastAPI sidecar, no HTTP shim, no two-runtime tax. LM-Kit links into your .NET process, picks up the right native acceleration, and stays out of the way. Async/await all the way down.

LM-Kit vs LangChain

Compare

vs. other .NET local stacks

Most ship inference only. LM-Kit ships the full runtime: agents, RAG, OCR, structured extraction, speech, vision, classifiers, embeddings, plus the symbolic layer that makes small models behave.

LM-Kit vs LlamaSharp

Workflows where local AI actually wins.

LM-Kit.NET is built for the .NET applications that cannot send data to a cloud endpoint, cannot rely on a network connection, or cannot afford per-token costs at scale.

Regulated

Healthcare, finance, government

Designed for environments with HIPAA, GDPR and data-residency obligations. Patient records, claims, and citizen data stay on the box.

Enterprise

Internal copilots

RAG over policies, runbooks, wikis, contracts, support tickets. Cited answers without sending source material to a third party.

Edge

Offline and air-gapped

Field laptops, rugged kiosks, vehicle telemetry, manufacturing floors. Inference works without connectivity.

Cost

High-volume workloads

Batch document processing, classification pipelines, customer support analysis. Marginal cost is compute, not token bills.

Product

Shipping AI in a desktop app

Wrap LM-Kit in a Windows/macOS desktop product. Customers run inference on their own hardware. No backend to operate.

Speed

Latency-critical UIs

Voice assistants, code editors, interactive analysis. Local inference removes the round-trip; first-token times in milliseconds.

Zero dependencies. One NuGet.

Add a single package to your .csproj. The runtime, native binaries for every supported backend, and the entire AI stack come with it. No Python runtime, no Docker, no daemons.

terminal 
# 1. Add LM-Kit.NET to your project
$ dotnet add package LM-Kit.NET

# 2. (Optional) Plug in a GPU backend
#    The dependency package is pulled in transitively.
$ dotnet add package LM-Kit.NET.Backend.Cuda13.Windows
# or:
$ dotnet add package LM-Kit.NET.Backend.Cuda13.Linux
Program.cs 
using LMKit.Model;
using LMKit.TextGeneration;

var model = LM.LoadFromModelID("qwen3.5:4b");
var chat  = new MultiTurnConversation(model);

var reply = await chat.SubmitAsync("Hello, LM-Kit.");
Console.WriteLine(reply.Text);
Open the Quickstart API reference Code samples GitHub samples

Free for small teams. Commercial above the line.

One license, both products, no activation key. Nothing checks a license at runtime. Evaluation and development are free at any company size, with no time limit.

Free

$0no key, no expiry

The complete SDK and the complete server, including commercial use and redistribution, for small companies. Evaluation and development stay free at any size.

  • Under $1M USD annual gross revenue
  • 10 or fewer employees
  • No more than $3M USD raised from outside investors
  • Always free: personal, education, nonprofits, open source
Download

Professional

Customannual, scaled to scope

Required above the thresholds. Scaled to deployment size, never metered by tokens, seats or end users. Carries the assurance a production deployment needs.

  • Commercial use and redistribution at any scale
  • Long-term support builds and security patches
  • Support with response-time commitments
  • Unlimited developers and end users
Get a quote

Questions we hear often.

Does LM-Kit require a Python runtime?

No. Everything runs in-process inside your .NET application. No Python, no Docker, no daemons, no HTTP service. One NuGet, one process.

Which GPUs work out of the box?

NVIDIA via CUDA 12/13, Apple Silicon via Metal, AMD and Intel via Vulkan. CPU and AVX2 act as a fallback. The same C# code dispatches to whichever backend is fastest on the host.

Can I run on .NET Framework?

Yes. LM-Kit targets .NET Standard 2.0, so it slots into .NET Framework 4.6.2+ codebases alongside .NET 8 / 9 / 10.

How do I add a model that is not in the catalog?

Point new LM(new Uri("...")) at any GGUF on Hugging Face or your own storage. The catalog is a curated set, not a constraint.

Does data ever leave the box?

Never, unless you explicitly call an external tool like WebSearch. Model inference, RAG, OCR, speech, and embeddings all run on your hardware.

Can I fine-tune a model?

Yes. LM-Kit ships in-process LoRA training and full fine-tuning. The same NuGet that runs inference runs the training loop.

What about Microsoft.Extensions.AI and Semantic Kernel?

Both are first-class bridges. Existing IChatClient pipelines and SK connectors work unchanged with LM-Kit as the local backend.

Is there an MCP client?

Yes. The agent runtime includes Model Context Protocol clients. Built-in tools and external MCP tools coexist in the same registry under the same permission policy.

Ship private document intelligence this sprint.

Install the NuGet, load a model, ship the feature. No key, no signup, and the free tier covers the entire surface.

Quickstart View on NuGet Pricing Talk to us