Brevitas

Brevitas Systems develops tools that make AI models’ working memory more efficient to manage. Its product, Splice, edits a model’s key-value (KV) cache—the stored state used to generate subsequent tokens—in place, rather than discarding it and rebuilding the context from scratch. Splice can delete spans of text, replace them with spans of the same length, or reorder content, including information from images, video and audio.

It adjusts positional data and repairs affected computations, then checks each edit against a fresh re-ingest. Edits that pass its correctness gate can be served; others fall back to a clean prefill, so speed is not prioritized over correctness. Brevitas reports edits up to 12.

4 times faster than re-prefilling, with larger gains on media workloads when encoder processing can be skipped. The company also offers the bvx CLI, benchmarks, documentation and research on cache editing and verification. Brevitas invites teams running self-hosted open models on SGLang or vLLM with frequently changing contexts to benchmark Splice on their workloads