Corbenic AI's Galahad, a KV-cache memory layer for self-hosted LLMs, now ships a free licence that covers one NVIDIA GPU for non-commercial use.
- Galahad keeps what a model has already read and replays it byte-exact instead of recomputing it, plugging into vLLM, SGLang and llama.cpp.
- Free licence:
pip install galahad-kv, thengalahad free --org <key> --email <email> --accept-non-commercial. It is bound to a single NVIDIA GPU and needs Linux x86-64 with Python 3.10–3.14. - Numbers from the project: 99.6% of tokens served from memory, 14× faster on vLLM, and GPU memory flat at 34.1 GB while reading texts up to 50 million tokens.
- Paid licences are priced per server, from $59.22/month (RTX 4080 SUPER) to $1,423.50/month (B300 SXM) on the rate card.