Local AI decision utility
Plan the workbench before you buy the hardware.
Estimate the memory headroom, runtime architecture, and hardware tier for your real local-AI workload.
This is an explainable planning estimate, not a benchmark, compatibility guarantee, or product recommendation sponsored by a manufacturer.
How the estimate works
The planner uses an editorial estimate: parameter count × selected bytes per parameter × 1.08 for approximate weight-format overhead, then adds fixed headroom of about 3GB for gateway-only work, 5GB for local inference, or 8GB for a combined gateway/UI stack, multiplied by the selected concurrency. These are planning assumptions, not measured runtime requirements. Exact usage varies by model architecture, runtime, context length, offloading, quantization format, and whether memory is unified or dedicated.
Lily is a specialized Apple silicon engine, not a general runtime.
Perplexity’s official blog describes Lily as a lightweight local inference engine built specifically for Apple silicon and Qwen3.6-35B-A3B so on-device compute does not bottleneck Hybrid Compute in Perplexity Computer. Official facts: Rust runtime, OpenAI-compatible API, custom Metal kernels, neither PyTorch nor MLX in the execution path, standalone demo on GitHub. It is not a directory SKU and not a 4.5 product rating. Do not treat it as a drop-in replacement for Ollama or LM Studio.