Today, SAGE Magnus completes its Phase 1 pre-training. As we move into alignment and reinforcement learning, we want to address the public interest this project has generated and clarify what Magnus is, and how it was built.
An MoE Architecture
Magnus is a Sparse Mixture of Experts model, not a dense one. This distinction matters. Training a large dense model from scratch was not computationally viable for our team. MoE allowed us to maintain high reasoning capacity while keeping active parameter counts tractable during both training and inference. This architectural choice was deliberate, not incidental.
From Validation to Pre-training
Magnus is not a fine-tune of an existing model. It is a ground-up pre-training run.
The path to it involved significant prior work. Over the past year, we validated our Meta-Cognitive Head (MCH) and Inverse Reasoning (IR) mechanisms by applying them via SFT on open-weight models including Qwen and LLaMA. Once we confirmed that MCH and IR reliably produced stable multi-step reasoning behavior, we integrated them directly into the base architecture rather than bolting them on afterward.
How We Secured Compute
The most common question we receive, particularly in Nepal, is whether this is real given our resource constraints.
It is. We do not have hyperscaler backing. We secured GPU clusters through infrastructure partnerships, combined with aggressive MoE routing optimization to minimize wasted compute. We are not in a position to name all partners at this stage, but we will share more as those relationships develop.
The Dataset
Our early SFT work used distilled outputs from frontier models to rapidly validate architectural hypotheses. That approach served its purpose for experimentation. It was not appropriate for a foundational pre-training run, for both practical and licensing reasons, and we did not use it here.
The Magnus pre-training corpus is built primarily on permissively licensed open-source datasets including Fineweb, Dolma, and localized public domain sources. Our focus was on aggressive filtering, deduplication, and internal synthetic verification using our own models rather than volume for its own sake.
Next Steps
Skepticism is reasonable and we welcome it. Over the coming months Magnus will go through iterative alignment before any external deployment. We will continue sharing updates as the work progresses.