Gemma Tuner Multimodal vs SMF (Semantic Memory Filesystem)

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.

SMF (Semantic Memory Filesystem) is an open-source Python library that treats the POSIX filesystem as the native memory infrastructure for AI agents. The core bet: instead of standing up a vector database, embedding service, and retrieval pipeline, you model your agent's memory as ordinary directories, files, and symlinks — then use the OS's own tools for retrieval. Entities are directories, relationships are symlinks, metadata is file attributes, and search is built on grep and find. The appeal is radical simplicity. Every developer already understands the filesystem. Memory built on top of it is inspectable with any editor, versionable with git, and portable across machines with rsync. There's no new query language to learn, no vector index to maintain, and no external service to keep running. Dynamis-Labs argues that for many agent memory use cases, semantic similarity search is overkill — you need entity graphs and efficient lookup, which the filesystem already provides. With only 7 stars and created yesterday (April 14), SMF is in very early stages. But the approach has attracted immediate discussion from developers frustrated with the operational overhead of vector databases for relatively structured memory tasks. It's a contrarian bet that's worth watching.