Belgian startup launches beta tool to save AI attention states to local storage, enabling servers to skip redundant text processing.
Key points
- Galahad beta launched October 1 2026, saving encrypted KV cache to local NVMe drives.
- The system supports vLLM, SGLang, and Llama.cpp engines for exact state reuse.
- Corbenic AI credits a 14x speedup primarily to Blaise retrieval, not just storage speed.
Belgian startup Corbenic AI launched the beta version of Galahad on October 1 2026. The tool saves encrypted key-value caches to local NVMe drives, allowing servers to reuse attention states without rereading text.
What happened
Corbenic AI introduced Galahad as a method to prevent GPUs from processing identical text multiple times. Founder Sietse Schelpe told TechRadar Pro that the system requires no special hardware beyond the NVMe drive already present in the server. The beta release entered on October 1 2026.
Galahad saves a model’s key-value cache, which represents the attention state built while reading a prompt. This data is stored in encrypted local storage and reloads when the same input bytes appear again. The system works with vLLM, SGLang, and Llama.cpp engines.
The company claims a 14x speedup, but its paper attributes most of this gain to Blaise, a retrieval component that reduces the amount of text read. Schelpe noted that the stored cache only activates after GPU memory clears or the server restarts.
Why it matters
Large language models consume significant memory to process prompts. Storing these states on disk allows systems to bypass recomputation when identical inputs recur. This approach extends the utility of existing hardware without requiring additional GPU memory upgrades. Teams can leverage local storage to maintain performance under high load conditions.
The fix
- Verify that your server has compatible NVMe storage available for cache operations.
- Ensure encryption is enabled for all stored key-value cache data to protect sensitive inputs.
- Monitor GPU memory usage to understand when cache eviction triggers disk storage use.
Why it matters
Organizations should evaluate how their AI workloads handle repeated inputs. Understanding the difference between in-memory reuse and persistent storage helps plan capacity. Check whether your current inference engines support external cache integration. Assess the security implications of storing processed data on local drives. Ensure that retrieval mechanisms do not introduce latency that outweighs the benefits of skipping computation.
What to do and how to stay safe: Corbenic AI
- Audit current GPU memory usage to identify bottlenecks caused by repeated prompt processing.
- Review storage encryption standards to ensure cached attention states are protected at rest.
- Test retrieval components in non-production environments to measure actual speedup benefits.
- Consult vendor documentation for compatibility with your specific inference engine versions.
Step-by-step guide: What is password policies and how they protect systems
General security guidance from the Crypt Weekly newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Does Galahad require new hardware to function?
No, founder Sietse Schelpe stated that the system uses the NVMe drive already present in the server.
Which inference engines does Galahad support?
The beta supports vLLM, SGLang, and Llama.cpp engines for key-value cache reuse.
What causes the 14x speedup claimed by Corbenic AI?
The company's paper attributes most of the speedup to Blaise, a retrieval component that reduces text reading.
