ReClaim runs entirely in your environment. Your platform team hosts our model weights and stands up the MOSES codec in front of them; after that, your applications redirect to an OpenAI-compatible endpoint of your own. The codec translates every prompt into MOSES — the language the model natively speaks — so it reaches the same answer over roughly 30% fewer tokens, and no prompt ever leaves your infrastructure.
Your platform team does the deployment once. After that, every application in your organisation reaches it the same way it reaches any OpenAI-compatible endpoint — by pointing at a URL.
We provide the ReClaim base model weights. Your team serves them on your own GPUs, in your own environment, under your own controls.
A Python service runs the MOSES codec and exposes an OpenAI-compatible API in front of the model.
From here it is a base-URL change. Your code, SDKs, schemas, and prompts stay exactly as they are, and every app reaches the same endpoint.
The MOSES codec translates your English prompt into MOSES — the language our model natively speaks — then translates the reply back. In between, the model works over roughly 30% fewer tokens, and the response comes back in the format your application expects.
The ReClaim base model is built on the Qwen family and natively speaks MOSES, the codec’s language. Known provenance, known licensing — and the weights are delivered to you.
The front end exposes standard request and response shapes. If your application already talks to an OpenAI-compatible endpoint, it already talks to this one.
The codec and the model both run on your infrastructure. Prompts and responses never transit a ReClaim service, because there isn’t one in the path.
Applications reach ReClaim through a base URL, so pointing them back is a one-line change. Nothing is embedded in your application code to unwind.
The base model is efficient as delivered. The gains that matter to your business come from training on what your business actually knows.
A Qwen-derived foundation that natively speaks MOSES, delivered as weights you host yourself alongside the Python codec front end.
Train on your own data and attach the result to the ReClaim base model. Your domain knowledge, our efficiency — without retraining a foundation model from scratch.
The economics read differently depending on where you sit.
Every agent, retry, and RAG call adds token volume. MOSES removes roughly a third of it before the model ever sees the prompt — and your teams keep the applications they already shipped.
See the capacity model →Deploy ReClaim inside your own platform and offer it to your customers. They get an efficient model with their data attached; you get a differentiated AI tier running on infrastructure you already operate.
Partner with us →Everything runs inside your perimeter — no prompts sent to a third party, no external inference call to justify. The codec is deterministic and lossless, so the same prompt always behaves the same way and results stay reproducible under audit.
Read the security brief →Deploy ReClaim in a staging environment and send real traffic through it. We’ll help you stand it up and measure token reduction, latency, and throughput on your own infrastructure — with your own data.