The MSE Graph Language Model (MSE-GLM) is a deterministic and explainable language modeling architecture that eschews neural network weights in favor of a graph-based approach. Instead of relying on probabilistic methods and learned parameters, it represents language as a directed graph where tokens serve as nodes and observed transitions between tokens are edges. This structure allows every generation decision to be traced back to explicit, inspectable rules, offering full transparency into the model’s inference process.
Designed for domains where guarantees, reproducibility, and auditability are paramount, MSE-GLM is positioned for use cases such as grammar-constrained generation, embedded AI, and tooling that requires an audit trail. It is not intended for open-domain generation or reasoning tasks typically addressed by transformer models. The architecture ensures that only valid, observed token transitions are generated, making it suitable for settings with a well-defined, finite output space, such as generating valid SQL clauses, JSON keys, or assembly mnemonics.
The model’s architecture is built around three compact, array-backed matrices: the Edge Matrix (E) that captures bigram relationships, the Bridge Matrix (B) for trigram context, and the Relationship Matrix (R), which enables lineage-aware tie-breaking and batch auditing of training data provenance. The inference engine operates through a four-stage pipeline, providing step-by-step explainability for each generated token. Ambiguities in possible next tokens are resolved using principled, inspectable rules rather than random sampling, ensuring deterministic output.
Training is performed in a single O(N) pass over the corpus, without backpropagation, epochs, or the need for a GPU. The tokenizer is a custom, from-scratch Byte Pair Encoding (BPE) implementation, supporting streaming training from files and preserving sentence boundaries. The resulting trained model is stored as a set of JSON files, which can be loaded and queried on any machine with Python, and the system is CPU-only.
MSE Graph Language Model is a Foundation models & chat product. It focuses on providing a transparent, deterministic alternative to black-box neural language models for explainable AI. MSE Graph Language Model is an open-source project aimed at AI researchers and developers seeking explainable language models. The project is open source (Open Source). It runs on the command line, and it can be self-hosted.
It is developed by Clifford Chivhanga, and the product first shipped in 2026. The project is developed in the open on GitHub with 6 commits in the last 90 days. Among its 7 catalogued features are deterministic inference, explainable outputs, and zero learned weights.
Latest indexed changes and source events
Zero Weights Language Model (MSE-GLM) verified by the PulseGate indexer
Other apps tracked under the same category.
