Backboard.io says new AI model compression cuts GPU needs by up to 70%

Backboard.io unveiled a set of AI products developed in Ottawa, including model compression, a coding suite, a multi-model chat app and a memory system it says ranks first on two benchmarks. The company is positioning the stack as a way for enterprises to run advanced AI with less hardware, lower cost and more data control.

Why it matters: - Backboard.io is targeting three persistent AI pain points at once: rising compute costs, data exposure to outside providers and the spread of unapproved employee tools. - The company says its approach can let organizations run advanced AI with less hardware, which matters as GPU supply, data-center capacity and power access remain constrained. - The full stack is designed for customers that need AI to stay inside their own environment, including governments, hospitals, banks and critical infrastructure operators.

What happened: - Backboard.io announced four products and capabilities developed in Nepean, Ontario. - The company said the work was done by a team made entirely of graduates of Canadian universities, colleges and CEGEPs. - The announcement includes BackboardQuant, Backboard Studio, Nash and an AI memory system. - The products are built to support enterprise use, sovereign deployment and consumer access in some cases.

The details: - BackboardQuant compresses AI models by up to 70%. - Internal testing showed compressed models retaining accuracy comparable to full-precision versions while running up to 2.7 times faster. - Backboard said the result can let one GPU handle the workload of two or three. - BackboardQuant ships built into enterprise deployments. - Backboard Studio matches coding tools from major AI labs on public benchmarks at up to 90% lower cost. - On Terminal-Bench 2.1, Backboard Studio scored 79.8% running Claude Opus 4.8. - Backboard cited published results of 78.2% for GPT-5.5 and 78.9% for Opus 4.8 on the same public harness. - Running the open-source GLM 5.2, Backboard Studio scored above 72% without a proprietary model. - A built-in token optimizer reduces frontier model usage by up to 30% on like-to-like comparisons. - Backboard Studio runs in the cloud or self-hosted, so source code does not leave the customer environment. - Nash gives users access to thousands of AI models across text and images in a single chat app. - Nash keeps user memory separate from the models. - Backboard says Nash is meant to reduce shadow AI by replacing unapproved tools with one sanctioned application. - Nash is available now for consumer and enterprise use at hellonash.ai. - Backboard says its AI memory ranks first on LoCoMo and LongMemEval, two independent benchmarks. - The company says the benchmark results are public and reproducible. - Memory stays inside the customer environment and under the customer’s control. - Backboard's full stack, including the API, application layer and models, can run inside a customer's own cloud. - The company says data remains in place and the system operates under the customer's governance. - Model compression lowers the hardware required to run advanced AI in place, extending deployment from large data centers to on-premise and edge environments.

Between the lines: - Backboard is framing AI memory, context and user data as the long-term value layer, not the model itself. - The strategy contrasts with the broader industry focus on building larger models and more AI hardware. - Sovereign deployment is the main sales pitch for regulated and security-sensitive buyers who want to keep data from moving to outside providers. - The benchmark claims are central to the company’s positioning, but they are only useful if customers can reproduce similar results in their own workloads.

What’s next: - Backboard says Backboard Studio and Nash are available now. - The company is pushing enterprise deployment of BackboardQuant and its memory stack inside customer-owned environments. - The next test is whether customers adopt the products as a lower-cost alternative to larger frontier-model setups.

The bottom line: - Backboard.io is betting that better memory, tighter deployment control and model compression can matter more than bigger models for many enterprise AI users.