A research paper on LoRA (Low-Rank Adaptation) concluded that the technique exhibits a trade-off where it "learns less but forgets less," meaning it is less prone to catastrophic forgetting but also absorbs less new information compared to full fine-tuning.

The open-source model GPT-OSS is considered the first open-weight model trained specifically with tool use in mind.

Open-weight models from China are popular partly because they have unrestricted open licenses, unlike models like Llama or Gemma which have usage restrictions for large companies.