Publication summary: GLM EXL3 eight-image acceptance¶
Canonical facts¶
Local functional acceptance: dual RTX PRO 6000 Blackwell Max-Q, pinned 4-bpw EXL3, FP8 DS-MLA KV, TP2/EP2/DCP2, speculation off, configured 327,680 tokens/C4, eight images/request, no video. The current recipe reconstructs r9. Four eight-card requests, one high-resolution eight-image request passed. The large request contained 63,456 input tokens and completed in 28.17 seconds. Six direct preflight checks passed. No speed ranking, broad visual-quality comparison or simultaneous eight-image soak is claimed. The prior one-image profile is the rollback.
Model: brandonmusic/GLM-5.3-Flash-tr3-4bpw@a5fee929cf4888b1824323e33e8a19b60129e025. Image: sha256:0f1cdcc8891f1cc3a444121eb61d366289a1cbba285f0892dcbb24bc94961692. The inherited runtime build metadata and its attestation limitation are in the finding. Direct and routed online requests followed cold startup, then warmed kernels and mixed image-cache state. Manifest · Evidence index.
X / short post¶
Local GLM EXL3 on dual RTX PRO 6000 Max-Q passed five direct eight-image tests. One 63K-token image batch took 28.17s. No concurrency-soak claim.
https://fakoli.github.io/anvil-serving/findings/2026-09-18-glm53-vision/
Reddit¶
Title: GLM EXL3 on dual RTX PRO 6000 Max-Q: eight-image functional acceptance
Pinned 4-bpw EXL3 with FP8 KV and speculation disabled retained its configured 327K window. Four eight-card comparison requests passed in forward/reverse order. Eight 2520×2520 cards also passed: 63,456 input tokens, 28.17 seconds for that one request. The retained Pi output matches all eight cards, but its request attachments and route are not independently correlated in this bundle. Coding/JSON/tool regressions passed. The limit is eight images/request, no video. This is synthetic functional acceptance, not a general performance or vision-quality ranking; simultaneous eight-image requests were not soaked. Full method, earlier negative results and raw evidence are in the linked finding.
Screenshot alt text¶
GLM EXL3 on two RTX PRO 6000 Max-Q GPUs passes four ordinary eight-image requests, one high-resolution eight-image request. Configured 327K context, FP8 KV, speculation off; no concurrent eight-image soak.
Claim ledger¶
| Claim | Conditions | Evidence |
|---|---|---|
| Four ordinary eight-image requests passed | Forward/reverse order, two repetitions, C1 | Native result |
| Large eight-image request passed in 28.17 seconds | One request, 2520×2520 each, 63,456 actual prompt tokens | Native result |
| Pi output matches eight cards | Output only; attachment count, route and served identity lack correlated provenance | Pi response |
| Six direct preflight checks passed | Includes four tool calls | Preflight |
| Eight/request; 327K configured | C4 scheduler, no simultaneous-eight-image soak | Recipe, limits |
The earlier one-image recipe passed 10/10 synthetic image attempts. Its 259,922-token retrieval refused once and passed on retry after reported overlapping traffic; causation is unproven. Those results remain separate dated evidence, not a fresh long-context gate for r9.