Models & hardware

Model choice, memory, storage, and compute shape the installation you operate.

Pre-release introduction. Public downloads are not available here yet. Assistant naming is planned; other capabilities depend on the version you obtain.

Plan around the actual model and workload. The documented primary model-serving target is Linux with NVIDIA CUDA-capable GPU nodes; CPU-only profiles serve bounded development and smoke workloads. Other engines and accelerators need a compatible, assessed profile.

Include room for stored knowledge, indexes, selected apps, backups, and concurrent work. Model licenses remain separate from Swan's software license. Remotely hosted machines and external services introduce their own operators into the trust boundary.