Brevitas Systems
Brevitas Systems provides Splice, a KV-cache surgery platform that makes large language model working memory editable in place during inference. The technology enables direct, certified modifications to the KV-cache, allowing developers to patch errors, inject new facts, or steer model behavior in real time without retraining or fine-tuning.
- Artificial Intelligence
- Developer Tools
- Software Only
Funding
Founders
Founder details are not available yet.
Product
Problem
Large language models rely on a KV-cache as their working memory during inference, but this memory is static and cannot be directly modified once generated. This inflexibility forces developers to retrain or fine-tune models, or use complex external memory systems, to correct errors or update knowledge in deployed models.
Solution
Brevitas Systems provides Splice, a KV-cache surgery platform that makes a model's working memory editable in place. The technology enables direct, surgical modifications to the KV-cache during inference, allowing developers to patch errors, inject new facts, or steer model behavior in real time. Splice certifies that these edits are correct, ensuring the modified model maintains its expected behavior and does not deviate from its intended outputs.
Target Audience
AI engineering teams and machine learning researchers working with large language models who need to update model behavior post-deployment without incurring retraining costs.
Features
- Direct in-place editing of KV-cache memory during model inference
- Certified correctness for all memory modifications to guarantee output integrity
- Real-time memory updates without requiring model retraining or fine-tuning
- Enables error correction, knowledge injection, and behavioral steering in deployed models