Blog
')"><br>Pioneer
GLiNER
Discord
')">
Back
How to fine-tune open weights models: Model selection, data curation, fine-tuning strategies, and evals
How to fine-tune open weights models: Model selection, data curation, fine-tuning strategies, and evals
August 5, 2026
August 5, 2026
Research
Open-weights models have seen a period of unusually rapid development this summer. New releases have arrived at a near-weekly pace, with several reaching performance levels comparable to leading closed models on independent benchmarks.<br>This has significant implications for users who want to reduce AI costs and ultimately own their own models. Open weights models can be downloaded and fine-tuned for specific tasks and domains, giving organizations full control over the resulting weights rather than dependence on a third-party API.<br>In this post, we’ll go over the difference between open-weights and open-source models, explain what LLM fine-tuning is, offer advice on when to fine-tune (and when not to), discuss some of the most popular open weights models available for fine-tuning, and talk about fine-tuning strategies, including generating data, choosing fine-tuning techniques, and evaluating your fine-tuned model.<br>Open-weights vs. open-source models<br>Not all LLMs can be fine-tuned by the public, and the dividing line is access to the model weights. Proprietary, or "closed," models like the Claude and GPT families can't be fine-tuned in the traditional sense because their weights aren't publicly available. Open-weights and open-source models can be, since both release the parameters needed to modify them. While the two terms are often used interchangeably, they describe different degrees of openness in how a model is released.<br>An open-source model is one whose weights, training dataset, and training code are released under a license that lets anyone freely use, study, modify, and redistribute the model. Nemotron 3 Ultra by NVIDIA is a recent example of a fully open-source family of models, released under the new OpenMDW-1.1 license. OpenMDW stands for open model, data, and weights.<br>An open-weights model refers to a model whose weights are released to the public, but without the underlying training data or training code. The permitted usage of open-weights models can vary. For instance, GLM 5.2 is an open-weights model released under the MIT license, making it free for anyone to download and run, with self-hosting unrestricted for commercial use. Kimi K3 is also open weights, but released under a more restrictive license that attaches revenue-triggered conditions for commercial operators rather than the unrestricted terms of MIT.<br>For the purposes of this article, we’ll focus on open-weights models.<br>What is LLM fine-tuning?<br>Fine-tuning is the process of taking a pre-trained language model and continuing to train it on a smaller, targeted dataset so that it adapts to a specific domain, task, or style. This adjusts the model's existing weights rather than training from scratch, which lets it specialize while retaining the general capabilities it learned during pre-training.<br>Fine-tuning open weights models offers several advantages over relying on a general-purpose API:<br>Better task performance. A fine-tuned smaller model can match or beat larger generalist models on the specific task it was trained for, since it specializes rather than spreading capacity across everything.
Lower latency and cost. Fine-tuning lets a smaller model reach the needed quality, and smaller models run faster and cheaper per call than large frontier ones.
Customization. The model adapts to your domain, terminology, formatting conventions, and tone rather than defaulting to generic outputs.
Data control and privacy. With open weights, fine-tuning can happen on self-hosted infrastructure, keeping sensitive training data in-house.
Ownership. The resulting weights are yours to deploy, version, and reuse without depending on a provider's API or pricing.
Choosing between fine-tuning, RAG, and prompt engineering<br>Fine-tuning is not the only option for adapting a language model's behavior to a specific task or domain without training a model from scratch. Lower-lift options include retrieval augmented generation (RAG), which supplies the model with relevant external knowledge at inference time, and prompt engineering, which shapes the model's instructions and examples in the input to steer its outputs.<br>Here are some guidelines for choosing between fine-tuning, RAG, and prompt engineering:<br>Use prompt engineering when you can get the behavior you want from your LLM simply by changing the instructions, examples, or structure of the input alone. It's the fastest and cheapest option to iterate on and is usually the right first step for most tasks.
Use RAG when your LLM needs access to specific, external, or frequently changing knowledge that isn't in its training data and when you need responses grounded in citations.
Use fine-tuning when you need...