Galahad is a software layer measured on a "per-server" basis, deployed between the user's workloads and GPU hardware, used to eliminate redundant computation in model inference. Its core idea is that certain computations already completed by a model can be remembered at the byte level with exactness and directly stitched together and reused in new requests, without recomputing tokens.
In terms of main capabilities, Galahad includes two engines: Merlin and Taliesin, which form a reuse loop. Merlin is responsible for determining which computations can be reused and verifying their safety. It hashes each incoming prompt and queries whether it has been processed before. The query is performed by a live registry in VRAM with a response time of less than two microseconds; there is also a persistent disk ledger that survives restarts and migrates between machines in a byte-identical way. Only when two independent hashes match is it considered a hit, in order to avoid returning incorrect memories. Merlin can also run independently as a content-aware deduplication tool, stripping redundant context before content reaches the model, including in pure API scenarios. Taliesin is responsible for stitching the state found by Merlin into live requests, achieving zero-token recomputation. It is described as producing recovered logits that are byte-for-byte identical to recomputed results, with equal SHA-256 and KL divergence of 0, and has been verified across two model scales and two GPU targets. Blocks are merged through position rebinding and recomputation fusion, enabling the model to reason across blocks, not merely read them.
In terms of usage, Galahad is a software layer positioned between the user's workloads and silicon. Users continue to rent servers anywhere, and Galahad licenses only that software layer. Its reuse mechanism is model-agnostic, any model can run within the framework, model weights and data remain 100% user-owned, and the official statement says it will not see user data or weights.
The applicable scenarios are mainly server environments that need to run model inference at scale and have a large amount of duplicate or similar computation, as well as scenarios where users want to reduce inference costs without changing models or cloud providers. It is officially said to "end paying twice" by reducing computational overhead through verifiable reuse.
In terms of limitations, the available information currently does not provide pricing, a specific list of supported models, deployment details, complete compatibility notes for various hardware platforms, or independent third-party benchmarks or a detailed technical paper. Whether it is suitable for all workloads, and the actual savings, need to be measured and verified based on the user's own workload.
