SOMA

About SOMA

Soma is a research-driven platform operating as a Bittensor subnet (subnet 114) focused on context compression for AI agents. It addresses the inefficiency and high costs associated with AI agents repeatedly sending large context histories to language models during complex tasks. Soma runs open competitions where miners develop and submit compression algorithms, which are evaluated on real agent workloads for both task quality and token savings. The best solutions are integrated into Soma’s production infrastructure, enabling developers to reduce inference costs without sacrificing agent performance. Soma offers an open API, browser tools, and open-source components, with plans for enterprise deployments and transparent benchmarking.