Clef.

Not a conversationalist but a decisive specialist: ask for a choice, score, classification, or extraction and receive a typed answer meant for a system to act on.

View on Hugging Face See LocalEngine variants ↓

Model makerCloudflare

Supported in LocalEnginemacOS

Upstream licenseApache-2.0

SourceClef model card ↗

Where it earns its space

Built for a real local rhythm.

These are good-fit starting points—not promises of benchmark performance. Choose the smallest listed option that comfortably covers your work.

Classification and routing

Structured extraction from text or images

API decision endpoints

In LocalEngine

Configured variants, by device.

The formats below are the downloads currently represented in LocalEngine’s model catalogs. GGUF runs through the llama.cpp-based runtime; MLX options use the Apple Silicon MLX runtime.

macOS

Clef Flash 9B · Clef 27B

Formats: GGUF Q4_K_M + Q8_0 vision projector
Local note: Flash fits Macs with 16 GB memory for one-pass decisions; the 27B model targets 32 GB+ for higher accuracy. Neither model is for chat.

About the source

Model identity, task category, supported formats, and license are based on the linked Hugging Face Clef model card and the LocalEngine macOS/iOS catalogs. Hardware guidance reflects the packaged variants, not a universal system requirement.