The release that turns HomeBred-LLM into a truly universal local LLM desktop app.
We’re excited to announce HomeBred-LLM v1.1.0 — the biggest update yet. This release brings full GGUF v2/v3 model support with the real llama.cpp engine, local model import (bring your own .gguf file — no HuggingFace required), and LoRA adapter attach, scale, and toggle controls. All while keeping the app a single self-contained binary: no Ollama, no Docker, no Python, no terminal.
What’s New in v1.1.0
GGUF v2/v3 Model Support (llama.cpp engine)
The core of this release. HomeBred-LLM now speaks two model formats:
• GGUF — the industry-standard llama.cpp format, powered by LLamaSharp 0.27.0 (native llama.cpp bindings). Works with LLaMA 3.x, Mistral, Mixtral, Qwen 2/2.5, Phi-3/4, Gemma 2/3, DeepSeek, and any other architecture llama.cpp supports.
• ONNX — existing support is fully preserved through ONNX Runtime GenAI.
An internal Inference Service Router picks the right engine for each model automatically, so the workflow feels identical regardless of format.
Import Local Models (no HuggingFace needed)
Have a GGUF file on your hard drive? Import it directly with the new “📂 Import local GGUF…” button in the Model Library. The app validates the file against its pure C# GGUF reader, extracts metadata (architecture, quantization, parameter count, context length) instantly, and registers it in your library — ready to start in one click.
Smart import diagnostics: Sometimes a GGUF file looks right but won’t load. HomeBred-LLM now detects two of the most common problems at import time — HuggingFace-style tensor names and tokenizer vocabulary size mismatches — and surfaces them as a ⚠ warning badge right on the model card, so you know before you even try.
LoRA Adapters: Attach, Scale, Toggle
Specialize any GGUF model with pre-trained LoRA adapters:
• Attach one or more .gguf LoRA adapters per model
• Set a per-adapter blend scale (1.0 = full effect)
• Toggle on/off without removing the adapter
• Adapters are applied when the model starts — Stop and Start to pick up changes
Note: LoRA adapters work with GGUF models only. ONNX Runtime GenAI doesn’t expose adapter APIs, so the Config screen shows a clear informational message for ONNX models instead.
Under-the-hood robustness
• Native log capture — the app now taps llama.cpp’s native log output and folds it into clear, actionable load-failure messages
• Pre-flight validation — tensor naming, vocab size, and attention dimensions are checked before loading, converting cryptic native errors into plain English
• Automatic database migration — existing installations upgrade in place, no manual steps
Installation
1. Go to the Releases page on GitHub – https://github.com/Eisenmann/HomeBred-LLM/releases
🛠️ Build from source
Requires .NET 10 SDK:
git clone https://github.com/Eisenmann/HomeBred-LLM.git
cd HomeBred-LLM
dotnet run –project HomeBred-LLM
dotnet publish HomeBred-LLM -c Release -r win-x64 –self-contained -o ./publish/win
dotnet publish HomeBred-LLM -c Release -r linux-x64 –self-contained -o ./publish/linux
dotnet publish HomeBred-LLM -c Release -r osx-x64 –self-contained -o ./publish/mac
GPU note: Both the CUDA 12 and CPU backends are bundled. LLamaSharp selects CUDA automatically when NVIDIA drivers are present and falls back to CPU otherwise.
Full Changelog
Added
• GGUF v2/v3 model support via LLamaSharp 0.27.0 / llama.cpp
• Inference Service Router — automatic engine dispatch between ONNX and GGUF per model
• Import local GGUF files — “📂 Import local GGUF…” button in the Model Library
• LoRA adapter management — attach, scale, and toggle .gguf adapters per model
• Import-time diagnostics — ⚠ warning badges for HF-style tensor names and vocab size mismatches
• Pure C# GGUF reader — fast metadata extraction without loading tensor weights
• Native llama.cpp log capture — clear, readable load-failure messages
Changed
• LLamaSharp bumped 0.21.0 → 0.27.0 (adds Qwen3 architecture support)
• Model Library empty state updated to mention local import
• Automatic SQLite schema migration for existing databases (Format, Warnings, LoraAdapters)
Fixed
• Clear error messages for invalid GGUF files instead of opaque native exceptions
Supported Models
Any model available as a GGUF file works — LLaMA 3 / 3.1 / 3.2 / 3.3, Mistral / Mixtral, Qwen 2 / 2.5, Phi-3 / Phi-4, Gemma 2 / 3, DeepSeek, and more.
Recommended quantizations: Q4_K_M, Q5_K_M, Q6_K
What’s Next?
We’re already working on what’s ahead for HomeBred-LLM. Stay tuned for:
• Analytics export to CSV / JSON
• Session search and message export
• Multi-GPU model splitting
HomeBred-LLM is free and open source (MIT). If you enjoy the project, please ⭐ it on GitHub and share your feedback!
Leave a Reply