Maximizing AI Efficiency: Key Considerations for Fine-Tuning and Self-Hosting Your LLMs

In today’s fast-paced AI landscape, developers face critical decisions when it comes to optimizing LLMs. A big one is determining when to fine-tune and when to self-host LLM models as each option can significantly impact cost, data security, and performance.

Fine-tuning allows developers to customize LLMs for specific tasks, enhancing their overall accuracy and relevance. Self-hosting, however, offers greater control over data privacy and resource management. In this post, we’ll delve into the essential factors and best practices that developers should consider to effectively harness the power of fine-tuning and self-hosting in their AI projects.

Fundamentals of LLM Fine-Tuning Techniques

In a recent workshop given at the Intel® AI DevSummit 2025, Yuri Winche Achermann, co-founder and head of software, data and operations at Tool Detective, walked developers through the nuances of fine-tuning LLMs (large language models) to improve their performance for specific tasks, emphasizing the importance configuring the environment for optimal performance on Intel's hardware.

Achermann explained how to load pre-trained models from Hugging Face*, test their initial performance, and then fine-tune them using specific datasets. He also showed how to use the LoRA (Low-Rank Adaptation) method to compress models for more efficient training, highlighting the importance of setting appropriate parameters to achieve desired outcomes.

"Fine-tuning is inherently iterative,” said Achermann, “we must continuously adjust parameters and evaluate model performance."

If you missed the workshop, check out the full recording and follow along with Achermann, an Intel Software Innovator and oneAPI Student Ambassador, as he guides you through how to set up an account on Intel® Tiber™ AI Cloud and run fine-tuning processes on a Gemma model. Then, go ahead and submit your fine-tuned models on Intel’s hardware to the Powered-by-Intel LLM Leaderboard  for evaluation and comparison. For additional insights and techniques on getting the most out of your LLMs, check out our top five tips for LLM fine-tuning and inference.

Navigating LLM Deployment: Tips, Tricks, and Techniques

In her talk on Navigating LLM Deployment: Tips, Tricks, and Techniques at the Intel® AI DevSummit 2025, Meryem Arik, co-founder and CEO of TitanML, shared practical strategies and best practices with AI developers on how to deploy LLMs within enterprise environments.

Understanding deployment boundaries to optimize model selection and performance is critical, according to Arik. Self-hosting involves the process of running models and GPUs in-house, either on-premises or in a virtual private cloud/cloud account. Arik emphasized the importance of determining when self-hosting is appropriate, highlighting cost efficiency, improved performance, and enhanced privacy and security as key benefits.

"Selecting the right models for specific tasks and consolidating infrastructure are crucial steps to maximize efficiency,” said Arik, who shared several best practices for self-hosting, including:

• Use quantized models to optimize resource usage
• Implement effective batching strategies to improve GPU utilization
• Leverage workload-specific optimizations

If you're looking for practical guidance on how to effectively deploy and manage LLMs, watch the full video recording here. By following Arik’s tips and techniques, you'll be better equipped to achieve cost-effective, high-performance, and secure AI deployments.

"Selecting the right models for specific tasks and consolidating infrastructure are crucial steps to maximize efficiency."

- Meryem Arik, Co-founder/CEO, TitanML

Hungry for More AI Knowledge?

Dive into more AI sessions from the Intel® AI DevSummit 2025 to learn from industry experts, explore the latest advancements, pick up best practices, and take your projects to the next level.

We encourage you to also check out and incorporate Intel’s other AI/ML Framework optimizations and tools into your AI workflow and learn about the unified, open, standards-based oneAPI programming model that forms the foundation of Intel’s AI Software Portfolio to help you prepare, build, deploy, and scale your AI solutions.