Lizard LLM

A hardware-aware local LLM runtime for Windows

Lizard LLM scans the CPU, available memory, and graphics hardware before recommending a compatible GGUF model. This removes the guesswork from choosing a model size or quantization for a Windows PC.

Two native inference engines

Lizard Native is a resident-GPU Direct3D 12 engine for models that fit on a compatible graphics card. Caterpillar uses a precompiled compute-graph approach suited to CPUs and shared or integrated graphics. Lizard selects an engine automatically and includes llama.cpp or optional Ollama fallbacks for unsupported cases. Explore the layers and provider comparison in the Technical overview.

Private local AI after setup

After the selected model is downloaded, prompts, responses, and model weights stay on the Windows device. Users can chat in the application or connect compatible software to the local OpenAI-style API.

What is included

  • Windows 10 and 11 support on 64-bit PCs
  • Hardware scan and compatible model recommendations
  • Local model downloads and integrity checks
  • Native CPU and GPU inference
  • Local chat and OpenAI-compatible API
  • On-device benchmark tools

Continue with the technical documentation, review the benchmark method, or download Lizard LLM for Windows.

Native-provider benchmark evidence

The product view includes the same 19-image evidence gallery as the benchmark page. It shows Lizard Native and Caterpillar as distinct local inference providers beside stock llama.cpp and optional Ollama, with warm-up, failures, throughput, native decode and missing rows left visible.

Runtime, throughput and native decode charts for llama.cpp Q4, Lizard Native Q4 and Caterpillar Q4