v2.0.0 - Major Architecture Overhaul (Physics Engine & Live Sync)

🚀 LocalAI Advisor v2.0.0 — Major Architecture Overhaul

Welcome to the v2.0.0 release of LocalAI Advisor! This major update transitions the recommendation and performance models from static heuristic tables into a physics-grounded inference speed estimator (inspired by llmfit), while adding live Hugging Face intelligence, multi-GPU support, and advanced catalog filtering.

---

✨ What's New in v2.0.0

• 🎯 Physics-Grounded Inference Speed Estimator (llmfit approach):

  • Decode throughput is now calculated directly from memory bandwidth: tok/s = (Bandwidth × Efficiency) / (Model Size + KV Cache).
  • Built-in memory bandwidth database for ~40 modern GPU models (RTX 50/40/30/20 series, AMD RX 7000/6000, Intel Arc).
  • Amdahl's Law layer offload latency modeling with PCIe transfer overhead for hybrid GPU+RAM execution.
  • Mixture of Experts (MoE) support: memory footprint based on total parameters, token rate based on active routing experts.

• 🎛️ Multi-GPU & RAM Bandwidth Detection:

  • Automatically aggregates VRAM across all dedicated GPUs in your system (total_dedicated_vram_gb).
  • WMI physical memory inspection detects RAM clock speed (MHz), memory type (DDR4/DDR5), and memory channels to compute exact system RAM throughput for CPU modes.

• 🌐 Live Hugging Face Auto-Sync & Enrichment:

  • Automatic background sync upon launch if cache is older than 30 minutes.
  • Real-time compatibility verification against your system's hardware limits.
  • Model release year extraction displayed on model cards (e.g. DeepSeek R1 Distill Qwen 14B (2025)).
  • Live "Last synced from Hugging Face: X min ago" status indicator in the status bar.

• 🔍 Advanced Catalog Filtering:

  • Toggle switch to filter / show uncensored and abliterated models.
  • Date range filter dropdown (All Time, Last 1 Month, Last 6 Months, Last 1 Year).

• 📐 Responsive Desktop Interface:

  • Disk cards and featured model cards dynamically adapt between 2, 3, and 4 columns based on window width.
  • In-app automatic update notification when a new GitHub release is published.

---

📦 Standalone Executable Download

Asset	Size	Platform
LocalAI-Advisor.exe	~19.44 MB	Windows 10 / 11 (64-bit)

---

🔒 SHA256 Integrity Verification

Verify your download using PowerShell:

Get-FileHash .\LocalAI-Advisor.exe -Algorithm SHA256

Expected SHA256:

A1FC334F7EDD96D729BDF78417F623562484D196A36E4CC525F17FCA00106D29

v1.3.0 - Initial Public Release

🚀 LocalAI Advisor v1.3.0 - Initial Public Release

Welcome to the initial public release of LocalAI Advisor, the standalone Windows diagnostic and local AI model recommendation engine!

---

✨ Highlights & Features

• ⚡ Zero-Overhead Hardware Diagnostics:

  • In-process detection of CPU instruction sets (AVX, AVX2, AVX-512, NEON).
  • Dedicated GPU VRAM detection (NVIDIA NVML / AMD / Intel) with 32-bit overflow protection.
  • Multi-threaded hardware scanning with zero UI freezing.

• 💽 Smart Multi-Disk NVMe/SSD Storage Engine:

  • Automatically queries Windows Storage WMI (MSFT_PhysicalDisk & MSFT_Partition) to identify physical drive media types (NVMe SSD, SATA SSD, HDD).
  • Prioritizes high-speed NVMe drives for 30x faster model load times into RAM/VRAM.

• 🌐 Live Hugging Face & Ollama Sync:

  • Dynamic discovery and parsing of trending GGUF models.
  • Robust parser supporting Mixture of Experts (MoE), decimal parameters (0.5B to 70.6B), and offline caching.

• 🎯 Precision Compatibility Scoring & Token Speed Predictor:

  • Evaluates memory footprints across Q4_K_M, Q5_K_M, and Q8_0 quantizations.
  • Categorizes top picks for ⚡ Ultra Fast, ⚖️ Sweet Spot, 🧠 Deep Reasoning (R1 / CoT), and 💻 Coding & Dev.

• 📋 1-Click Deployment & Built-in Ollama Guide:

  • Instant copy of ollama run <model> commands.
  • Step-by-step setup walkthrough and multi-drive storage path tips.

• 📦 Standalone Portable Executable:

  • No Python, dependencies, or environment configuration needed.

---

📦 Download & Run

Download LocalAI-Advisor.exe from the Assets below and double-click to launch on Windows 10/11 (64-bit).