GLM-5.3-Flash K3/DFlash2 optimization artifacts

Sanitized local evidence for the dual-RTX-PRO-6000-Max-Q campaign finalized on 2026-08-30. The authoritative narrative is the dated finding. These files contain synthetic benchmark prompts and outputs, loopback endpoints, and the exact served identity; they contain no credentials or reachable private network identity.

Publication redaction: Operator-specific absolute corpus paths in the two linked image-only artifacts were replaced with the stable repository-relative corpus path. Corpus identity and content hash, measurements, model outputs, and event ordering are unchanged.

Exact identity

• Target: wrldsuksgo2mars/GLM-5.3-Flash-EXL3-K3-v1@319d66a8b53092b491f698440ecea781e4ddd4e4
• Draft: incoai/GLM-5.3-Flash-DFlash2@dc77ff1c99eeb2df044ee3d4f0094eb033fee410
• Runtime image: sha256:001a45bd71bcf908a8c07459570bdb8c5e0a205d085f29ac7f3201529fa3eb75
• Hardware: 2x RTX PRO 6000 Blackwell Max-Q, 96 GB each, TP=2/DCP=2, PCIe without NVLink, Docker Desktop/WSL2

Research and feasibility

• External priors
• Preload feasibility input
• Preload feasibility result

The feasibility calculation classified the selected K3 target plus DFlash2 draft as a benchmark survivor at the required 250K+8,192/C2 envelope. It did not claim runtime compatibility or quality; those are established by the artifacts below.

Incumbent baseline

• Full functional preflight
• Bounded quality
• Image-only corpus
• c1 depth sweep: 4K, 32K, 65K, 131K, 240K
• short concurrency: C2, C4, C8, C16
• long C2: 131K, 240K

Selected K5/2048 profile

• Full functional preflight including 250K target
• Exact 950K-target retrieval
• Bounded quality
• Image-only corpus
• Restored direct acceptance
• Authenticated routed acceptance
• Rejected forced-thinking-off control
• c1 depth sweep: 4K, 32K, 65K, 131K, 240K, 500K
• short concurrency: C2, C4, C8, C16 run 1, C16 run 2
• long C2: 131K, 240K, 500K

Alternate and rejected settings

• K3 alternate: 4K c1, C16 run 1, C16 run 2, bounded preflight
• K5 with 4,096-token scheduler chunks, rejected: 4K c1, C16

Integrity and publication

• SHA-256 manifest
• Publication summary

The machine-readable benchmark artifacts are the measurement authority. The finding and publication summary are derivative views. External priors are not local results, and the client/routing promotion evidence is retained privately because it contains real operator topology.