TurboFieldfare on a Mac

Gemma 4 26B-A4B on Apple Silicon in ~2 GB of RAM · AI / LLMs

TurboFieldfare is a custom Swift 6.2 + Metal 4 runtime that runs Google's Gemma 4 26B-A4B (instruction-tuned) on any Apple Silicon Mac; including 8 GB machines; in roughly 2 GB of RAM, by streaming the mixture-of-experts weights from disk instead of holding them resident. Useful when you want a 26B-class local model on a Mac that can't hold one in memory; for other Gemma sizes use the Ollama gemma3:* tags in Model Manager instead, since this runtime serves only the one model. Built from source: git clone, swift build -c release, then run .build/release/TurboFieldfareMac (the ~14.3 GB model streams down on first launch). TurboFieldfareCLI gives instruction chat and raw completions with --temperature/--top-k/--max-new. An experimental loopback OpenAI-compatible Chat Completions server on http://127.0.0.1:8080/v1 adds streaming and function-tool support; add that in AI Models as a network endpoint to use it in the assistant. Apache-2.0. Note it builds unsigned third-party source rather than installing a notarized binary, and :8080 is also llama.cpp's default; change one if you run both.

Run TurboFieldfare with FrontierStack

Install or connect TurboFieldfare; then check its status, ports and certificate from FrontierStack. The same screen links to firewall checks, Malware Audit and backups where they apply.

Related in AI / LLMs

• Vane (Perplexica)
• MLX
• Needle 2
• Cerebras
• Darkbloom
• MLX (Apple)
• vLLM
• llama.cpp Server