DeepMind introduces SkillSmith to synthesize model weights natively

Treating model weights as a native LLM modality could fundamentally transform model customization by making weight synthesis as fluid as text generation. By unifying textual knowledge and parametric adapter weights into a single reasoning process, SkillSmith uses specialized KV adapters to ingest and emit prefix-tuning weights directly. This eliminates reliance on separate fine-tuning pipelines or manual weight-merging, laying the groundwork for autonomous, self-improving AI agents capable of synthesizing their own specialized skills.