The unofficial GLM-5.3 E224 derivative publishes weights, but its two-DGX-Spark target remains untested.
Unofficial GLM-5.3 derivative targets two DGX Spark systems
2026-10-07Verified announcementIGNORE
An unofficial derivative prunes GLM-5.3-Flash from 288 to 224 routed experts and quantizes routed experts to NVFP4, producing a 141 GiB weight footprint intended for two DGX Spark systems. The card reports benchmarks from a B200, not the target Spark hardware, and explicitly labels Spark throughput and memory behavior as estimates.
What this means for you
For most teams, this is a narrow, expensive deployment recipe whose central target claim has not been measured on the target hardware. High community interest does not justify engineering investment until independent two-Spark results and quality regressions are available.
Evidence & related material
Technical details & scores
| Derivation | unofficial GLM-5.3-Flash derivative; 224 of 288 routed experts retained; NVFP4 routed experts |
|---|---|
| Footprint | 151.5 GB on disk / 141 GiB weight memory; target is two 128 GB DGX Spark systems |
| Reported Validation | accuracy and throughput measured on one B200 |
| Target Limit | two-Spark memory and throughput are estimated, not measured |
| Community Signal | 5,422 downloads and 274 likes at review time |
Technical impact 2
Engineering value 1
Adoption signal 3
Maturity 2
Verification cost 5
Risk 4