Galahad-KV Enables 50M-Token Context via NVMe Cache
October 9, 2026
The galahad-kv package enables byte-exact KV state loading from encrypted local NVMe disk for Gemma 4 models. Testing on 50 million tokens shows loading is up to 4.3x faster and uses up to 12.3x less GPU energy than recomputation while maintaining flat GPU memory usage.
HOW THIS AFFECTS YOU
●
builderYou can implement massive-scale long-term memory for agents using local NVMe storage instead of expensive recomputation.