Galahad Saves AI Working State to SSD, but Its Speed Claim Needs Context
Corbenic’s Galahad beta stores reusable model state on local SSDs. TechRadar’s review of its paper shows that retrieval drives the biggest reported speedup.
The Short Version
Corbenic AI’s Galahad beta saves part of a model’s working state to encrypted local storage so it can reuse earlier processing of the same text, TechRadar reports. It targets work that would otherwise need repeating after GPU memory is cleared or a server restarts. The approach is concrete, but its roughly fourteenfold speed claim needs qualification.
Why It Matters
The saved material is a key-value, or KV, cache: state produced as a model processes a prompt. Galahad’s Taliesin component identifies stored blocks using the exact input, model and customer identity. In principle, loading matching state can avoid computing it again.
A second component, Blaise, instead selects a relevant section of a document for the model to read. That is a different source of savings: doing less work, rather than simply restoring previous work.
The Bigger Picture
According to TechRadar’s account of Corbenic’s paper, adding Blaise raised the reported llama.cpp speed gain from 3.1 times to 14.5 times. Taliesin alone was slower than the baseline in the paper’s vLLM and SGLang tests, while improving speed with llama.cpp.
Those results also changed what the model could see. The baseline’s limited context covered only the final 12% of the test corpus and answered 10 of 100 questions; Taliesin could attend to all of it and answered 98. Neither timing nor accuracy should be read as a comparison with identical conditions.
What to Watch
TechRadar reports that the founder ran the paper’s experiments himself on rented cloud GPUs. Raw logs were promised but unpublished, and Blaise’s selection method was withheld pending patents. Independent tests with equivalent context, accuracy requirements and competing cache systems would help establish where the beta is useful. The reported figures do not establish a general fourteenfold speedup from SSD storage alone.
Would faster AI answers matter to you if the comparison also changed how much source material the model could see?
Connie Brew