v0.16.0 — BitNet Auto-Download & Native NuGet Packages

What's New

BitNet Auto-Download from HuggingFace

• GGUF models now auto-download on first run — zero manual setup for model files
• BitNetModelDownloader with cache-first logic via ElBruno.HuggingFace.Downloader
• BitNetChatClient.CreateAsync() factory with progress reporting
• EnsureModelDownloaded option (default: true)

Platform-Specific Native NuGet Packages

• ElBruno.LocalLLMs.BitNet.Native.win-x64 — Windows x64 native library (llama.dll)
• ElBruno.LocalLLMs.BitNet.Native.linux-x64 — Linux x64 native library (libllama.so)
• ElBruno.LocalLLMs.BitNet.Native.osx-arm64 — macOS ARM64 native library (libllama.dylib)
• NativeLibraryLoader probes runtimes/{rid}/native/ paths (NuGet convention)
• Improved error messages with platform-specific NuGet package suggestions

CI/CD

• build-bitnet-native.yml — cross-platform bitnet.cpp build workflow (3 runners)
• publish-bitnet-native.yml — native NuGet publish via OIDC trusted publishing

Tests

• 58 new BitNet tests (NativeLibraryLoaderTests + NativePackageValidationTests)
• Total: 229 BitNet tests passing

---

Full Changelog: v0.15.0...v0.16.0

v0.15.0 — BitNet 1.58-bit Model Support

What's New in v0.15.0

New Package: ElBruno.LocalLLMs.BitNet

Run Microsoft's BitNet 1.58-bit ternary models in .NET through IChatClient.

Install:

dotnet add package ElBruno.LocalLLMs.BitNet

Highlights:

• BitNetChatClient implementing IChatClient from Microsoft.Extensions.AI
• Wraps bitnet.cpp (llama.cpp fork with ternary kernels) via P/Invoke
• 5 model catalog: BitNet 0.7B, 2B-4T (default), 3B, Falcon3 1B, Falcon3 3B
• Models as small as 150 MB with excellent CPU performance
• Streaming, DI registration, platform-specific native lib loading
• 155 unit tests

Samples:

• BitNetChat - basic chat completion
• BitNetPerformance - benchmark BitNet vs ONNX models

Docs:

• BitNet Guide: docs/bitnet-guide.md
• Blog Post: docs/blog-bitnet-launch.md

v0.11.0 — RAG XML Docs + Comprehensive Tests

What's Changed

Fixed

• XML Documentation (Issue #12): Added comprehensive XML doc comments to all 13 public types in ElBruno.LocalLLMs.Rag, eliminating all 116 CS1591 warnings

Added

• 60 new unit tests (Issue #11): Comprehensive test coverage for the RAG package
  • RagRecordTests.cs (27 tests) — record types construction, equality, immutability
  • SqliteDocumentStoreTests.cs (14 tests) — SQLite persistence layer
  • RagServiceExtensionsTests.cs (13 tests) — DI registration
  • LocalRagPipelineConstructorTests.cs (6 tests) — constructor validation

Stats

• 0 build warnings (was 116)
• 813 total tests (718 xUnit + 95 MSTest), all pass
• ElBruno.LocalLLMs: 0.10.0 → 0.11.0
• ElBruno.LocalLLMs.Rag: 0.1.0 → 0.2.0

v0.10.0 — Zero-Cloud RAG Sample

What's New in v0.10.0

🔌 Zero-Cloud RAG Sample (Closes #9)

A complete offline RAG pipeline — no cloud APIs needed. Combines three ElBruno packages:

• ElBruno.LocalEmbeddings — real ONNX-based vector embeddings (all-MiniLM model)
• ElBruno.LocalLLMs — local LLM inference via Phi-3.5-mini-instruct
• ElBruno.LocalLLMs.Rag — chunking, storage, and retrieval pipeline

dotnet run --project src/samples/ZeroCloudRag

The sample loads documents, chunks them, generates embeddings, indexes in an in-memory vector store, retrieves relevant context for a query, and streams a grounded answer from the local LLM. 11 steps, zero cloud.

🧪 Tests

27 new tests across 3 files:

• 10 MSTest unit tests for LocalRagPipeline
• 4 MSTest E2E integration tests (gated behind RUN_INTEGRATION_TESTS)
• 13 xUnit tests for RAG record types

Total: 757 tests, all pass.

📚 Documentation

• docs/rag-guide.md — new Zero-Cloud RAG section with DI architecture
• README.md — ZeroCloudRag added to samples table
• docs/supported-models.md — RAG model recommendations

Full Changelog: v0.9.0...v0.10.0

v0.9.0 — Qwen2.5-Coder-7B ONNX + OpenAI Server

What's New in v0.9.0

🧑‍💻 Qwen2.5-Coder-7B-Instruct — Code Assistant Model

The first code-specialized model in the library! Converted to ONNX INT4 (6.3 GB) and published to elbruno/Qwen2.5-Coder-7B-Instruct-onnx on HuggingFace.

• Same Qwen chat template as the rest of the Qwen2.5 family
• Supports tool calling for agent-based coding workflows
• Works out of the box — HasNativeOnnx = true

var options = new LocalLLMsOptions { Model = KnownModels.Qwen25Coder_7BInstruct };
var client = new LocalChatClient(options);
var response = await client.GetResponseAsync("Write a quicksort in C#");

🌐 OpenAI-Compatible HTTP Server Sample

New src/samples/OpenAiServer — a minimal ASP.NET Core server that exposes local ONNX models via OpenAI-compatible REST endpoints.

• POST /v1/chat/completions (streaming SSE + non-streaming)
• GET /v1/models (list available models)
• Works with VS Code Copilot custom models, Continue, Cody, and any OpenAI SDK client
• Includes chatLanguageModels.json config for VS Code integration

📋 Blocked Models Documentation

• Codestral 22B v0.1 — MNPL-0.1 license prohibits production use
• Devstral Small 2 (24B) — no ONNX conversion path (custom Tekken tokenizer, FP8 quantization)

✅ Quality

• 705 tests pass (8 new for Qwen2.5-Coder model definition)
• 0 warnings, 0 errors across entire solution

Full Changelog: v0.8.0...v0.9.0

v0.7.2

What's Changed

• feat: DX improvements — Issue #7 fix + exceptions + logging + diagnostics + builder (#7) by @elbruno in #8

Full Changelog: v0.7.0...v0.7.2

v0.7.1

What's Changed

• fix: MaxSequenceLength reports effective runtime limit by @elbruno in #6

Full Changelog: v0.6.0...v0.7.1

v0.7.0

What's Changed

• fix: MaxSequenceLength reports effective runtime limit by @elbruno in #6

Full Changelog: v0.6.0...v0.7.0

v0.6.1

What's Changed

• feat: expose model metadata (context window, model name) via LocalChatClient by @elbruno in #4

New Contributors

• @elbruno made their first contribution in #4

Full Changelog: v0.5.0...v0.6.1

v0.6.0

What's Changed

• feat: expose model metadata (context window, model name) via LocalChatClient by @elbruno in #4

New Contributors

• @elbruno made their first contribution in #4

Full Changelog: v0.5.0...v0.6.0