Skales v12.7.45 - SKLS Unsloth
A repair release, built the day after 12.7.4 from what the first devices
We've just released Skales v12.7.45, a minor update focused on improving the stability and reliability of your local AI experience. This release addresses a number of issues reported by users, particularly around local server connectivity and model management.
Unsloth Desktop Integration
Unsloth Desktop is now a first-class provider within Skales. You'll find its own tile in the provider grid, complete with an on/off switch and an address field. A button is available to locate a running instance on its standard ports. As Unsloth Desktop is the first local runtime requiring an API key, the card now clearly indicates where to obtain it, and refused keys are reported correctly. The "Refresh Models" function now reads Unsloth Desktop's GGUF catalogue, showing quantisation details for each file. Model VRAM duration is now configurable within Unsloth Desktop itself.
Local Server Connectivity
A key fix in this release ensures that when you switch to Skales Local, the correct local model ID is used. Previously, the previous provider's model ID might have persisted, leading to requests for models that the local runtime couldn't find. Both request paths now correctly resolve model IDs from the capability matrix, provider card, or the runtime itself. Requests to local addresses will no longer carry foreign vendor IDs.
Mac Apple Silicon Build
The Mac build for Apple Silicon now correctly includes the local server. Previous versions were unnecessarily large, carrying engines for other platforms. Each platform and architecture now stages its own engines, and build gates ensure the correct files are included. Bundled libraries have also had their build-machine search paths repaired, leading to smaller Mac packages.
Model Discovery and Loading
We've improved how Skales discovers and displays available models. The "Fetch Available Models" function now correctly identifies models held by your local runtime, displaying them directly on the card without needing a button press. Projector companion files will no longer appear as answerable models. Additionally, models located on other disks are now found more reliably, with Skales checking Ollama's configured paths, LM Studio settings, and asking running runtimes what they hold.
New here? Skales is the private AI agent you install like an app - free for personal use, any model, everything stays on your machine.
Download Skales