Understanding models through architecture, training and evaluation.
We investigate efficient language-model design and the training systems behind it. Our OMN research notes describe what we have implemented, what internal analysis suggests and what remains unverified.
I. Architecture and optimization belong together.
OMN is our model architecture and training research programme. OMN-1 combines a dense decoder-only Transformer with parallel residual blocks, grouped-query attention, query/key normalization, rotary position embeddings and a Muon-plus-AdamW optimizer recipe.
These components build on existing research. Our work is to understand their interaction in a concrete implementation, improve the training system and test which choices contribute to useful capability.
II. Efficiency is a question to measure.
We want capable models that make careful use of parameters, training compute and data. These are research objectives, not established advantages over other architectures.
Meaningful comparisons need matched data, compute budgets and evaluation protocols. Smaller size, an unusual optimizer or a successful training run alone does not demonstrate superior reasoning or data efficiency.
III. Evaluation comes before capability claims.
Our internal OMN-1 records include pretraining analysis, weight-spectrum diagnostics, targeted ablations and context-length tests. They also document limitations.
We distinguish observations from explanations. A spectrum is not a measure of intelligence; one ablation cannot establish a universal mechanism. Configured context and tested context are reported separately, and no architecture is assumed to guarantee privacy or eliminate memorization.
IV. Specialization should meet engineering reality.
Our industrial-AI direction explores models designed for technical work and systems that connect them with operational context, calculations and engineering constraints.
We are interested in private deployment, evidence-grounded assistance and evaluation on defined tasks. Physical safety and operational permissions must remain explicit system responsibilities, not presumed properties of a language model.
V. Research should make its limits visible.
The notes below describe our current implementation and selected internal analyses. They are technical notes, not peer-reviewed papers or independently reproduced benchmark results.
As work progresses, we intend to report the methods, conditions and limitations needed to interpret findings. Model releases and industrial deployments will require their own evidence.
Research dispatches & audits
Three technical notes based on our OMN-1 implementation and internal post-pretraining records. These are not peer-reviewed papers; each note distinguishes observations, interpretation and limitations.
OMN-1: architecture and optimizer design
The dense decoder, parallel residual blocks and Muon–AdamW training recipe.
OMN-1: interpreting spectra through ablation
What recorded weight-spectrum and truncation experiments establish—and their limits.
OMN-1: configured context is not demonstrated context
Lessons from the recorded long-context probes and training-data inspection.