Fine-Tuning LLMs with Nvidia DGX Spark and Unsloth (docs.unsloth.ai)

Unsloth now supports local fine-tuning and on-device inference of very large language models on NVIDIA DGX Spark, claiming capability up to ~200B parameters on systems with 128 GB of unified memory. The project demonstrated practical results — gpt-oss-20b was fine-tuned with reinforcement learning on DGX Spark to reliably win the game 2048 (about 1,000 steps / 4 hours of RL gave a large performance boost), illustrating that substantial task-specific improvements are achievable on-prem without cloud services. This makes high-end DGX Spark boxes a viable path for organizations that need privacy, low-latency deployment, or cost control for large-model tuning and serving. Technically, Unsloth supplies Docker images and Jupyter notebooks (e.g., the gpt_oss_20B_RL_2048_Game.ipynb) to run QLoRA-style 4-bit fine-tuning workflows on DGX Spark; gpt-oss-120b QLoRA 4-bit training is reported to consume ~68 GB of unified memory, leaving headroom for larger experiments. The workflow supports RL-based fine-tuning and model checkpointing so you can save and deploy tuned models directly on the DGX Spark. For practitioners this means accessible, reproducible local pipelines (Docker/virtualenv + notebooks) for scaling fine-tuning to multi-billion-parameter models while leveraging DGX Spark’s unified memory and GPU fabric.