Unsloth fine-tunes LLMs on NVIDIA RTX GPUs

Unsloth now supports fine-tuning on NVIDIA RTX AI PCs and DGX Spark, enabling developers to customise models for tasks such as personal assistants, product support and long-context workflows.

The company said in a statement that Unsloth is optimised for efficient, low-memory training on NVIDIA GPUs, from GeForce RTX desktops and laptops to RTX PRO workstations and DGX Spark, and that the new Nemotron 3 family provides a starting point for agentic fine-tuning.

Unsloth supports parameter-efficient methods such as LoRA and QLoRA, full fine-tuning, and reinforcement learning. The company said parameter-efficient approaches typically require about 100–1,000 prompt-sample pairs, while full fine-tuning generally needs 1,000+ pairs. It added that reinforcement learning requires an action model, a reward model and an environment for the model to learn from.

The company said Unsloth can boost the Hugging Face transformers library performance by 2.5x on NVIDIA GPUs. NVIDIA also introduced the Nemotron 3 family in Nano, Super and Ultra sizes; Nemotron 3 Nano 30B-A3B is available now and is optimised for tasks such as debugging, summarisation and retrieval. The Nano model’s hybrid Mixture-of-Experts design delivers up to 60% fewer reasoning tokens and a one million-token context window, and Nemotron 3 Nano fine-tuning is available on Unsloth. Nemotron 3 Super and Ultra are expected to be available in the first half.

DGX Spark provides a compact, local option for memory-intensive fine-tuning, offering 128GB of unified CPU–GPU memory and FP4 performance up to a petaflop. The company said DGX Spark lets developers run larger models, including those above 30 billion parameters, use full fine-tuning and reinforcement-learning workflows, and avoid cloud queues for compute-heavy tasks.

The Recap

• Unsloth now supports fine-tuning on NVIDIA RTX hardware.
• Nemotron 3 Nano 30B-A3B available with one million-token context.
• DGX Spark offers 128GB unified memory for larger fine-tuning.