#3131st of 57 in Model serving
turbo-fieldfare
Swift and Metal runtime that runs Gemma 4 26B-A4B in about 2 GB RAM
- Stars
- 7.1k
- License
- Apache-2.0
- Last commit
- Oct 2026
- Last release
- Oct 2026
- Language
- Swift
0 on 18 Jul 20267,101 on 31 Oct 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
TurboFieldfare is a model-specific Swift and Metal inference runtime for Gemma 4 26B-A4B (4-bit MLX affine) on Apple Silicon. It keeps the 1.35 GB shared core and FP16 KV cache in memory and streams only the needed experts from SSD through a 16-slot LFU cache. It ships a native Mac app, a CLI, an experimental loopback OpenAI-compatible server and optional image input.
Who it is for: Mac users wanting a large local MoE model on 8 GB Apple Silicon
Strengths
- Runs a 26B MoE model in about 2 GB of weights plus 4K KV cache on 8 GB Macs
- Measured 5.1-6.3 tok/s on an 8 GB M2 Air and 31-35 tok/s on a 24 GB M5 Pro
- Includes Mac app, CLI and loopback server with Chat Completions, streaming and function tools
- Installer streams and repacks weights with hash-verified manifest, no second full checkpoint on disk
Weaknesses
- Supports only one pinned model, Gemma 4 26B-A4B instruction checkpoint
- Requires macOS 26, Metal 4 and Swift 6.2; arm64 only
- Needs about 14.3 GB of SSD space and a roughly 15 GB first download
- Server has no authentication or TLS and is marked experimental; audio and video unsupported
What it needs
- RAM ≥ 8 GB
- GPU required
- Needs Swift 6.2+, macOS 26 SDK, Metal 4, Hugging Face (model download)
- Models: Gemma 4 26B-A4B IT (MLX 4-bit, .gturbo format)
- port 8080
Quick answers
Is turbo-fieldfare open source?
Yes. Its code is on GitHub under the Apache-2.0 license.
Does turbo-fieldfare need a GPU?
Yes, it needs a GPU. Plan for at least 8 GB of RAM.
Also in Model serving
See all 57| Rank | Project | Score |
|---|---|---|
| 1 | llama.cppC/C++ inference engine serving GGUF models over an OpenAI-compatible API | 93 out of 100 |
| 2 | OllamaLocal model runner with CLI and REST API for open-weight LLMs | 91 out of 100 |
| 3 | LocalAISelf-hosted inference server with OpenAI-compatible API and pluggable backends | 90 out of 100 |
| #4 | vLLMInference engine for Hugging Face LLMs with PagedAttention and continuous batching | 88 out of 100 |
| #5 | colibriC inference engine that streams MoE experts from disk to run huge models | 85 out of 100 |