GLM mixed 3.5-bpw loader recovery: publication summary

The loader repair made the candidate runnable under a 60 GiB RAM limit. C1/C4 preflight and repeated bounded C4 quality checks passed. Retrieval succeeded at 216,307 actual input tokens. Strict capacity remained inconsistent, so the baseline was restored without promotion.

Copy

Short post (224 characters): GLM mixed 3.5-bpw now runs. C4 quality and 216K-token retrieval passed; strict capacity was inconsistent. Baseline restored, no promotion. https://fakoli.github.io/anvil-serving/findings/2026-09-17-glm53-mixed35-fix-forward/

Reddit title: Repaired GLM mixed 3.5-bpw loader: working C4, no promotion yet

Reddit body: A synchronous R7 weight-transfer patch removed CPU retention. The candidate passed bounded repeated quality and one 216,307-token retrieval. Its first strict C4 cell completed 4/4 at 98.17 completion tok/s, including reasoning; the next completed only 3/4. The baseline also failed strict output counts, so there is no valid throughput comparison or demonstrated overall win.

Screenshot alt text: Candidate and baseline test table separating quality, actual context tokens, valid capacity, strict failures, and no-promotion.

Claim ledger

Claim	Evidence
Loader recovery and contained OOM probe	Containment
C4 preflight and repeated bounded quality passed	Preflight, quality
Exact retrieval at 216,307 actual input tokens	Needle
One valid C4 cell, one failed repeat	First cell, second cell
Baseline restored, no promotion	Restoration, decision