← All Hot Projects

Hot Projects · Weekly archive

Audacity 4 and Slotstream: audio editing and local Qwen inference

This retrospective selection examines a dedicated Apple Silicon MoE runtime alongside ongoing structural transitions in foundational open-source audio tooling.

Week of Added September 8, 2026

audacity

Audacity 4.0, released on September 3, gives the cross-platform audio editor a rebuilt Qt interface and a revised clip-editing workflow. Users can select and edit several clips together, group clips, and save interface layouts as workspaces. The update suits podcasters and other creators who record and arrange audio, while existing users will need to learn where some controls have moved.

Compatibility is the main adoption question. The release notes list features still missing from 4.0, including Time Tracks, Note/MIDI tracks, Macro Manager and the scripting pipe. Projects use a new .aup4 format; opening an .aup3 project converts it without changing the original, but the converted project cannot be saved back as .aup3. The project is GPL-3.0 licensed, with separate terms identified for some components. Use the release notes to check a workflow before migrating it.

slotstream

Slotstream is an inference runtime written in Swift and Metal that serves the 125-billion-parameter Qwen3.8-Flash-Next 4-bit mixture-of-experts model on Apple Silicon Macs. Targeted at local AI developers and tool builders, it streams routed experts directly from disk into unified memory while keeping a 3.8 GB dense trunk resident. It implements the standard OpenAI and Ollama HTTP endpoints, allowing local applications, WebUI tools, and coding agents to interact with the model without requiring Python dependencies.

The tool imposes strict hardware and scope limits. It operates exclusively on Apple Silicon running macOS 14 or later, relies on unified memory architecture, and explicitly rules out discrete GPUs, Linux, or Windows support. Deployment demands roughly 110 GB of free SSD storage for its 105.3 GB weight pull, with performance heavily gated by disk read speeds and prompt prefill overhead capped at 32,768 tokens. In version 0.2.0, the Ollama CLI client fails to connect due to strict API validation, though curl and OpenAI SDKs function. Slotstream is licensed under the MIT license, with underlying model weights governed separately by the Qwen community license.

Sources