93%
Smaller than frontier models
120B vs ~1.8T parameters
6B
Active parameters per query
vs 37B+ for MoE systems
80–95%
Less energy per inference
vs general-purpose AI
5%
Parameters active at once
Sparse activation
The Problem
General-purpose AI has a
hidden cost
The largest AI models contain up to 1.8 trillion parameters. Routing every student submission through these systems carries a significant environmental footprint.
Research shows the most energy-intensive models consume
33+ Wh per long prompt, over 70× more than smaller, purpose-built alternatives.
Energy per long prompt (frontier model)
33+
Wh
70× higher than purpose-built alternatives
Model Architecture
A model built for one thing: assessment
Eduface Model
120B
total parameters
5%
Active at inference
6B
active per query
~5%
activation ratio
Sparse MoE architecture routes each task to only the relevant parameters
no wasted compute.
Frontier Models (est.)
~1.8T
total parameters (GPT-4, reported)
37%
Active at inference
37B+
active per query
15–20×
more compute
Designed for breadth across all knowledge domains, most of which are irrelevant to academic assessment.
DeepSeek activates
37B / 671B
5.5%, 6× more than Eduface
Eduface activates
6B / 120B
5% — domain-optimised minimum
How It Works
Sparse activation: only what is needed
Step 01
Submission arrives
A student's assessment is received by the Eduface platform.
Step 02
Router activates experts
Our MoE architecture routes the task to only the
6B relevant parameters.
Step 03
Minimal output
Feedback is generated using a fraction of the energy of a model.
For Universities
Sustainability is now a
procurement decision
Routing tens of thousands of student submissions through frontier AI
carries a cumulative energy cost that is not trivial. Choosing Eduface
means choosing infrastructure engineered to be sustainable from day one.
"Universities that invest in AI should invest in AI that was built for them. Domain-specific models are not only more accurate — they are significantly more sustainable."
E
Eduface
References & Sources
Fedus et al. (2022) Switch Transformers, JMLR. · Shazeer et al. (2017) Sparsely-Gated MoE, ICLR. · Jiang et al. (2024) Mixtral of Experts, arXiv:2401.04088. · DeepSeek AI (2024) arXiv:2412.19437.
· Caravaca et al. (2025) arXiv:2511.05597. · You et al. (2024) Scientific Reports 14, 26310. · Dettmers et al. (NeurIPS 2024) D2DMoE.
GPT-4 parameter count unconfirmed by OpenAI; figures reflect reported industry estimates. Eduface figures (120B total, 6B active) reflect our own architecture.