Updates
Technical notes, implementation findings, and verified milestones published directly by Aethr researchers and engineers.
OMN-1: architecture and optimizer design
OMN-1 is a dense, decoder-only Transformer implemented in PyTorch. Its documented 1.5B configuration combines parallel residual blocks, grouped-query attention, query/key RMS normalization, SwiGLU, rotary position embeddings and tied embeddings. This note separates the architecture from its optimizer and describes implementation rather than unproven performance advantages.
OMN-1: interpreting spectra through ablation
The internal audit inspected weight singular values and intervened on a selected attention projection. Removing low-magnitude directions increased evaluation perplexity in the recorded experiment. This supports testing functional effects rather than diagnosing a model from condition numbers alone.
OMN-1: configured context is not demonstrated context
OMN-1 has a 32,768-position configuration, but the recorded retrieval probes did not demonstrate reliable performance across that window. This note separates positional support, attention boundaries, training exposure and measured retrieval.