Gemma Tuner Multimodal vs Skrun

Gemma Tuner Multimodal is an open-source fine-tuning toolkit for Google's Gemma 4 and Gemma 3n models that runs entirely on Apple Silicon using PyTorch with Metal Performance Shaders (MPS) backend — no NVIDIA GPU or cloud infrastructure required. It supports LoRA training on multimodal inputs: audio, images, and text simultaneously, using local CSV files or streamed from Google Cloud Storage or BigQuery. The tool targets the growing segment of developers who own M-series Macs but have been locked out of fine-tuning workflows that assume CUDA availability. Gemma 4's architecture is particularly well-suited to this use case: its 4B multimodal variant (designed for on-device deployment) trains efficiently on M3 Max and M4 Pro hardware within the available unified memory constraints. Primary use cases include medical transcription fine-tuning (audio → text with clinical terminology), visual QA systems (image + text → structured response), and private on-device pipelines where cloud API calls are prohibited by compliance requirements. The project fills a specific niche that Google's own fine-tuning documentation doesn't cover well for Apple hardware.

Skrun is an open-source tool that wraps agentic skills — the discrete, reusable capabilities you build for AI agents (web search, data extraction, file transformation, API calls) — into deployable REST APIs with a single command. The idea is that skills you build for one agent context shouldn't be locked to that agent's runtime. With Skrun, you define a skill once with a standard function signature, and get a hosted endpoint with automatic request validation, retry logic, rate limiting, and an OpenAPI spec generated automatically. The project addresses a real architectural tension in the current AI tools ecosystem: agent skills are written in a dozen different formats (LangChain tools, MCP tools, function call JSON, OpenAI tool specs) and are essentially stranded assets — they only work within their specific orchestration framework. Skrun normalizes this by wrapping any skill definition format and exposing it as a framework-agnostic HTTP endpoint that any agent or pipeline can call. This appeared on Hacker News with a small but thoughtful discussion focused on the "skills as microservices" architectural pattern. Critics noted that adding HTTP round-trips to every tool call introduces latency; proponents argued that the composability and reusability benefits outweigh the cost. The early version focuses on stateless skills; stateful/conversational skill deployment is on the roadmap.