Model experiments
A smaller GPT-OSS artifact: the 28-expert prototype
A smaller model file is a useful engineering result. Preserving answer quality is a separate question.
- Experts retained
- 32 → 28
- Recorded storage reduction
- 1.19 GiB
The question
Can an architecture-aware expert reduction produce a smaller GGUF that still loads and returns useful smoke-test answers?
How the work was approached
A GGUF-aware script retained 28 of 32 experts using static router weight norm plus bias magnitude; metadata and affected tensors were rewritten under a new model name.
What the record shows
The historical report records a 1.19 GiB reduction and a successful LM Studio load. Reopened later raw run records show seven nonblank completed requests without visible reasoning/workflow markers.
What remains unresolved
Seven-request cleanliness checks are not semantic quality or noninferiority evidence. No current artifact acceptance or R640 benchmark. Some original model/plan paths no longer exist in current workspace; the original file cleanup is documented in the historical result.
What we are trying to achieve
The aim is to make a capable model easier to store and run on available hardware. The prototype establishes an engineering path for reducing an artifact’s size; the next important question is how much useful behavior survives under independent, held-out evaluation.
Sources & reading notes
- GPT_OSS_20B_PRUNING_EXPERIMENT.mdActual 28-Expert Prune Result Retained project record Source record retained privately.
- prune-gpt-oss-gguf-28x.pycompute_keep_plan static norm and bias scoring Retained project record Source record retained privately.
- report.mdHistorical seven-request run Retained project record Source record retained privately.
This is a dated, source-based research account, not a live machine report. Historical results, engineering preparation, research designs, open review gates, and measured pilots remain distinct.