turbo-fieldfare
TurboFieldfare’s README describes a custom Swift and Metal runtime designed to run the instruction-tuned Gemma 4 26B-A4B model within an approximate 2 GB memory footprint on Apple Silicon Macs. Written by Andrey Mikhaylov, the runtime targets developers seeking local text inference on hardware configurations with 8 GB of RAM. Rather than staging the full 14.3 GB quantized model in memory, the engine keeps a 1.35 GB core and KV cache resident while streaming routed experts from an SSD into a bounded cache during decode passes.
Adoption requires an arm64 Apple Silicon Mac running macOS 26, Metal 4, and Swift 6.2 or later, with around 14.3 GB of disk space for the installed .gturbo model layout. The package ships as an Apache 2.0-licensed project offering a native Mac interface, a command-line tool, and an experimental loopback OpenAI-compatible server. Scope remains restricted strictly to text generation from the pinned Gemma 4 26B-A4B checkpoint, without native support for audio, images, or direct tool execution outside loopback server declarations.
qm
The qm repository provides a self-hosted, multi-user agent framework tailored for team collaboration across web interfaces and Slack. Geared toward engineering teams and startups, it assigns each user or project channel an isolated workspace equipped with durable sandboxes, scoped memory, dedicated file stores, and credential sets. The platform connects interchangeable external execution harnesses—such as Pi, OpenCode, Codex, and Claude Code—to a shared headless core, so teams can choose among supported agents.
Running qm requires a TypeScript-compatible Node environment backed by PostgreSQL for session and state durability, with infrastructure targeting deployment across cloud providers like AWS or Fly.io. Scaffolding is managed via the @yc-software/qm package using the qm init command, which materializes deployment repositories without requiring modifications to the upstream core. The codebase is distributed under the MIT license, and operational security relies on configurable postures that dictate whether tool invocations execute automatically, require classifier screening, or pause for human approval.