Performances
RD Popularity Index
RD Growth Index
Comparison
Follow Updates
Follow Lists
People
About MoonMath
MoonMath, formerly Ingonyama, built its base accelerating zero-knowledge proving on GPUs and now applies that capability to private inference with the zro endpoint: developers use an OpenAI-compatible interface to reach open-weight models such as GLM, MiniMax and Kimi from coding agents including Claude Code, Codex CLI and Cursor. It addresses closed APIs that retain prompts and train on data, with requests neither stored nor added to training sets, inference running on its own GPUs, and the team maintaining AMD decode kernels and HyperQuant quantisation to lower unit cost.
It sits between inference infrastructure and cryptography engineering: unlike resellers of API access it owns hardware and kernels and can tune model, memory and operators together; against general cloud providers it sells non-retention and verifiability to development teams worried about leaking code and data. Model coverage is limited to open weights, closed frontier models are not included, and capacity is concentrated in data centres in the United Kingdom and the European Union.
Ingonyama focuses on ZK hardware acceleration. Key milestones include a $21M seed round, the March 2023 launch of the open-source GPU library ICICLE that lowered the barrier for ZK acceleration, and the subsequent v1.0.0 release with full developer support.
Long-term significance lies in the hardware roadmap: progressing from FPGA research to ASIC deployment. The integration of Accseal's LEO ASIC into ICICLE v3 allows the same library to support both GPUs and dedicated chips, while the ZaKi architecture aims to cut ZK cloud costs—marking a concrete step toward purpose-built ZK hardware.
In 2026 the former Ingonyama team rebranded as MoonMath AI, redirecting its zero-knowledge and GPU acceleration work toward large model inference. In June it published HyperQuant, a low-rank quantisation method that compresses key-value caches and weights and was also applied to video diffusion weights, and in the same month soft-launched a private inference service for coding agents.
In July MoonMath added GLM5.2 to the service, running on its own European GPUs with no prompt retention and no training on customer data; it also shipped hand-tuned kernels for AMD Instinct and upstreamed work to vLLM and SGLang. The product line includes acceleration components such as LiteAttention, LiteLinear and BackLite.
MoonMath will use the Zro endpoint to serve coding agents, stack in-house kernels, compression and hardware tuning into a cost advantage, and keep open-source components running alongside commercial service.
Omer Shlomovits — Founder & CEO; zero-knowledge researcher; previously founded the zk hardware acceleration company Ingonyama.
Yuval Domb — Chief Architect & VP; background in cryptography and chip architecture.
Danny Shterman — VP of R&D.