Lizard LLM

Lizard Native and Caterpillar technical architecture

Lizard separates hardware and GGUF model planning from inference execution. A supported model can run through Lizard Native or Caterpillar; bundled llama.cpp and optional Ollama remain explicit compatibility fallbacks.

Wireframe Lizard Native and Caterpillar inference providers converging on one local compute layer

Lizard Native

Lizard Native is a Windows-first Direct3D 12 provider. Supported quantized weights remain in resident GPU buffers while the model is active, compute runs through cached shaders, and model-fit planning uses the machine's available memory.

Lizard Native resident-weight Direct3D 12 architecture

Caterpillar

Caterpillar is a standalone clean-room implementation with its own GGUF reader, typed directed acyclic graph validation, lifetime-reused activation arena, and precompiled execution plan.

Caterpillar typed graph and reusable activation architecture

How the providers differ from llama.cpp and Ollama

Lizard's native providers are specialized execution lanes inside a hardware-aware Windows product. llama.cpp is a broad cross-platform inference toolkit and server. Ollama adds model packaging, management, scheduling, and APIs. Lizard uses those broader runtimes as measurable baselines or compatibility fallbacks, not as undisclosed native execution.

Continue to the provider-separated benchmark evidence, technical documentation, or Windows download.