GLM-5.3-Flash-9B-Surgery-Dummy-etheroi

GLM-5.3-Flash 9B Surgery Dummy

Experimental text-only test model. Not an official Z.ai release and not yet quality-recovered. Do not use it as a production or benchmark model.

This checkpoint preserves GLM-5.3-Flash width/kernel geometry while reducing the decoder from 45 to 10 layers and each routed MoE from 288 to 32 experts. It has 8,895,622,684 text parameters. Source layers are [0, 1, 2, 3, 8, 18, 25, 31, 38, 44]. Each target routed expert is a four-donor functional mosaic: 512 individually selected, coupled SwiGLU units come from each donor (gate/up rows plus matching down columns), a closed-form down-projection scale matches synthetic output variance, and all three expert matrices receive fresh per-128x128 FP8 E4M3 scales. Router rows use balanced router-space clusters. No donor forward pass, activation cache, distillation, or parameter training is used. Student-only evaluation remains required before calling the model functionally useful.

Vision is intentionally disabled. A zeroed 49,056-parameter visual compatibility stub exists only because the stock Transformers wrapper currently constructs a visual submodule. It is not a vision model.

The exact source revision, tensor map, expert clusters, and provenance hashes are stored in surgery_plan.json; output shard hashes are in surgery_manifest.json.