Transformer Lab

Transformer Lab is a cross-platform desktop application for the full local-LLM development loop. The application is open source under AGPL-3.0 at github.com/transformerlab/transformerlab-app, and is built with Python for the backend (around 67 percent of the codebase) and TypeScript for the desktop UI (around 28 percent). It runs as a native desktop app on macOS (Apple Silicon M1 through M4), Linux (NVIDIA or AMD GPUs), and Windows (NVIDIA via WSL2). The UI can also be detached from the engine: users can run the desktop interface locally while the actual training and inference engine runs on a remote GPU box or cloud machine.

The application packages multiple parts of the stack into one tool. For model management it supports HuggingFace, GGUF, and MLX formats with conversion between them, and one-click downloads for Llama, Gemma, Qwen, Phi, Mistral, and similar families. For training it supports full fine-tuning, LoRA, QLoRA, RLHF variants including DPO, ORPO, and SimPO, and explicit reward modeling. Hardware-specific trainers are provided for Apple Silicon (MLX), NVIDIA (CUDA), and AMD (ROCm). For inference it integrates MLX, vLLM, Llama.cpp, FastChat, and HuggingFace Transformers, with engine switching from the UI. A drag-and-drop dataset builder pulls from HuggingFace datasets or accepts local files, and a RAG mode handles retrieval over user files.

The project is credited to Ali Asaria and Tony Salomone, with Mozilla Builders backing. Mozilla selected Transformer Lab for the first Mozilla Builders Accelerator cohort, announced September 23, 2024, with awards of up to $100K plus mentorship under the cohort's local-AI theme. The most recent release as of mid-2026 is v0.38.0 (May 13, 2026).

The grant sits at the training, weights, and runtime layers of the open stack. The category it fits is the integrated local-LLM workbench, sibling to LM Studio, Oobabooga, and AnythingLLM, but distinguished by first-class training and fine-tuning support rather than inference-only chat. The AGPL license enforces source-availability for any networked redeployment, which is a stronger copyleft posture than most peers.