Skip to content

Publication summary: GLM-5.3-Flash APC

The finding and artifact set are authoritative. This copy is ready for review; nothing has been posted.

Canonical facts

Pinned GLM-5.3-Flash EXL3 4-bpw runs on two RTX PRO 6000 Blackwell Max-Q GPUs with v84, FP8 MLA KV, TP2/EP2/DCP2, configured 327680 context, C4 and no speculation. Shared-prefix mean visible TTFT fell 67.84%, request throughput rose 3.17 times, and actual cache reuse was 75.58%. Fresh-prefix request throughput fell 2.39%. Each finalist cell passed 32/32; final diagnostic quality passed 12/12 per arm. The exact baseline is restored. Independent review recommends bounded APC promotion after human approval; no production promotion has occurred.

X / short post

GLM APC on two RTX PRO 6000 Max-Q GPUs: shared-prefix visible latency 23.7→7.6s, 3.17× completed requests/s; fresh-prefix throughput −2.39%. Matched 32-request C4 cells, about 25K actual prompt tokens, strict output/canaries. Synthetic workload; not a coding-intelligence gain. Baseline restored; promotion still human-gated.

Reddit

We tested the same pinned GLM-5.3-Flash EXL3/v84 configuration with APC off/on. The warm shared-prefix population improved substantially while the fresh-prefix population showed a small cost. All four 32-request finalist cells passed exact output and canary checks, and isolated counters measured 75.58% cache reuse. Quality, images and long-context diagnostics passed. Reasoning volume differed, so request throughput and visible latency lead the report. This does not prove SWE quality or four simultaneous maximum-context sessions.

Screenshot alt text

Matched baseline and APC latency charts show a strong shared-prefix benefit and a small unique-prefix cost. Each cell contains 32 requests at C4.

Claim ledger

Claim Conditions Evidence
Shared TTFT −67.84%; completed requests 3.17× Warm shared cache, 32 requests each, C4 Derived comparison and native source hashes
Unique request throughput −2.39% Separate fresh-prefix 32-request cells Derived comparison
Cache reuse 75.58% Isolated before/after counter interval Cache delta
Final quality 12/12 each Three repetitions, bounded diagnostics Baseline, candidate
Baseline restored; no promotion Exact identity and direct/routed acceptance Restoration, review