Fastino publishes guide to fine-tuning open-weights models
Fastino has published a practical guide to fine-tuning open-weights models, arguing that a recent wave of capable releases makes owning task-specific models a realistic way to reduce AI costs.
Fastino has published a guide by Mary Newhauser on fine-tuning open-weights language models. It covers model selection, data curation, training strategies and evaluation.
The guide starts from what it calls an unusually rapid summer of open-weights releases. New models arrived almost weekly, and several reached performance comparable to leading closed models on independent benchmarks. The author says this lets organizations fine-tune models for specific tasks and keep full control of the resulting weights instead of depending on a third-party API.
The guide separates open-source models, which release weights, training data and code, from open-weights models, which release only the weights. Its examples include:
- Nemotron 3 Ultra (NVIDIA): fully open source under the new OpenMDW-1.1 license
- GLM 5.2: open weights under the MIT license, unrestricted for commercial self-hosting
- Kimi K3: open weights under a more restrictive license that adds revenue-triggered conditions for commercial operators
It recommends trying prompt engineering first. RAG is for external or frequently changing knowledge, and fine-tuning is for consistent output formats, specialized styles, or cases where a smaller model needs to match a larger one at lower latency and cost. For choosing a base model, it points to the target task, the use case and deployment constraints such as cloud versus on-device. It notes that the base model sets the ceiling on what fine-tuning can achieve.
The source material provided ends partway through the base-model section, so this note does not cover the guide's recommendations on data generation, fine-tuning techniques or evaluation.