LLMRadar
A local-first Windows model advisor and benchmark cockpit that detects hardware and AI runtimes, separates local, downloadable and cloud options, and recommends models for practical workflows.
Who this product is for
Windows users choosing local AI models for real hardware, privacy requirements and practical workloads.
The decision moment
A model decision depends on VRAM, RAM, runtime support, task fit and privacy instead of a generic leaderboard score.
Model choice should start with your actual PC.
Model selection is dominated by broad benchmarks and marketing. LLMRadar focuses on whether a model actually fits the user’s hardware, privacy needs and task.
Turn hardware and runtimes into a shortlist.
- Rust-based local hardware probe integrated with the Windows application
- Ollama, LM Studio and local AI-tool runtime detection
- Goal-based recommendations for coding, chat, reasoning, vision, memory and agents
- Deep German and English localization plus Setup and portable ZIP release artifacts
Start with the task. Then filter by reality.
- Describe the task and available hardware
- Apply privacy and provider constraints
- Compare suitable model options
- Use the recommendation as a routing decision, not an absolute score
Hardware-first model choice
It starts with your hardware, not a benchmark leaderboard.
- Separates installed, downloadable and cloud model options instead of mixing them together
- Recommendations are tied to detected hardware and practical goals
- Benchmarks the models you actually run while keeping installation and deletion outside its scope
Recommendations need more than one machine.
The Radar family has been exercised across four real Windows PCs: two desktops and two laptops, covering Windows 10 and Windows 11, including systems with and without dedicated GPUs. Release readiness is still verified per current build, so newer builds keep their own open gates instead of inheriting a blanket compatibility claim from earlier test rounds.
What the current beta has actually passed.
The 0.3.0-beta.4 changelog records 50/50 product and regression checks passing. GPT-OSS 20B GPU-load failure was reproduced, and the CPU/RAM fallback loaded, generated, measured and unloaded successfully. Beta-4 fit and runtime gates are checked.
What still blocks a public release.
- The current Windows packages are not code-signed
- Recommendations remain guidance tied to the local catalogue and detected hardware
- Windows 10, second-PC, AMD/Intel GPU, fresh-profile, full packaging/install/uninstall and remaining UI validation are still open release gates
- Final installer, Start-menu/Desktop shortcut and application-icon polish remains before public release
- Model installation and deletion remain deliberately outside the current scope
Choose a model from the hardware outward
Start with Can My PC Run Local AI?, then read Which Local LLM Can My PC Run? and How Much VRAM Do You Need for Local AI?.
0.3.0-beta.4 is a testable beta with 50/50 product and regression checks recorded as passing and successful CPU/RAM fallback validation after the reproduced GPT-OSS 20B GPU-load failure. It is not public-download-ready; Windows 10, second-PC, AMD/Intel GPU, fresh-profile, full packaging/install/uninstall and remaining UI validation are still open.