Research at InventDB.
Next to the platform we run three research programmes. One has already become part of the platform: it is the engine that serves the open model in InventDB SOAR. Another will take the platform to databases larger than one machine, and we plan to release it by the end of 2026. The third is long-term work on how a mind learns its world.
InventDB Sparkle
The inference engine we built to run open-weight language models on our own GPUs. Written in Rust with GPU code made for the NVIDIA H200, it is an engine an InventDB SOAR instance can select, and it serves Qwen3.8-27B today.
629 tokens a second for one conversation writing JSON, and faster than vLLM and SGLang at writing answers in almost every case we measured.
How Sparkle works and how it measuresInventDB Cluster
One InventDB database across many machines in three availability zones, with three copies of every piece of data, so a database can grow past what one machine holds and keep answering when a machine, or a whole zone, fails. We plan to release it by the end of 2026.
235,707 new orders a minute in TPC-C at 1,000 warehouses on thirty machines, and SQL over a billion rows.
How a cell works and what it measuredInventDB Brahma
Research toward a human-like intelligence that learns how its world works from its own experience, starting from a small, counted set of building blocks, and writes what it finds as rules a person can read. It uses no language model.
783 moments on average to predict a new room right four times in five, where a small neural network needed about 9,500.
Read the researchHow we report research.
If you work on inference engines, distributed databases, world models or how minds learn, we would like to compare notes. Write to us at contact@inventdb.com.