# drumih/turbo-fieldfare

Gemma 4 26B-A4B inference in ~2 GB of RAM on any M-series MacBook

Repository: https://github.com/drumih/turbo-fieldfare
Canonical: https://ross.abutalabs.com/products/turbo-fieldfare
Language: Swift
License: Apache-2.0
License Family: permissive
Topics: gemma, gpgpu, llm, local-ai, macos, metal, on-device-ai, swift, apple-silicon, gemma4, gemma4-26b-a4b, llm-inference, on-device-llm
Last push: 2026-08-26T13:31:14+00:00

## Health v2 (maintenance only)
Score: 80/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 99, release rhythm 99, longevity 3
- inputs: {"age_days": 47, "days_push": 7, "days_rel": 8, "gap_med": 1, "n_releases_24m": 14}
- flags: young
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 6397, forks 400 (observed 2026-08-28T04:09:42.989446+00:00)

## What it is
TurboFieldfare is a custom Swift + Metal runtime that runs Gemma 4 26B-A4B inference in roughly 2 GB of RAM on Apple Silicon Macs by streaming mixture-of-experts weights from SSD. It ships as a native Mac app, CLI, and local OpenAI-compatible server.

## Use cases
- run a 26B LLM locally on an 8 GB MacBook
- on-device chatbot inference without cloud APIs
- serve a local OpenAI-compatible LLM endpoint on macOS
- run Gemma 4 with minimal RAM usage
- experiment with Metal GPU kernels for LLM inference
- private offline text generation on Apple Silicon

## When to choose
- you have an M-series Mac with limited RAM and want large local LLM inference
- you want a model-specific optimized runtime rather than llama.cpp or MLX wrappers
- you want a native macOS app plus CLI and local server in one tool

## When to avoid
- you need multi-model or cross-platform support
- you don't have Apple Silicon or macOS 26+
- you need fine-tuning or training, not just inference
- you can't spare ~15 GB of disk for the repacked model

## Facets
- artifact type: application
- maturity: active
- function: llm-inference, gpu-computing, cli, gui
- domain: large-language-models, artificial-intelligence, machine-learning
- platform: cli
- tags: metal, apple-silicon, gemma, on-device-ai, mixture-of-experts, ssd-streaming, low-memory, local-llm, openai-compatible-server, macos, swift, gpu, desktop

## Member repositories
- drumih/turbo-fieldfare (main) score 80

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:09:42.989446+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T17:45:01.872614+00:00, confidence not recorded.
  - readme: https://github.com/drumih/turbo-fieldfare (fetched 2026-08-28T04:09:42.989446+00:00, sha 4ef4224d5d4e)
- Data as of 2026-08-30T08:39:29.467469+00:00.