Skip to content

Publication summary: GLM-5.3-Flash SGLang 393K promotion

This file is derivative publishing copy. The dated promotion finding and linked raw qualification artifacts are authoritative.

Canonical facts

  • Model identity: ormandj/GLM-5.3-Flash-W4A16-NVFP4-K32-Experts-FP8-WO@c3cbb9891b67c741bcbf6b176dd7af9265b069db, served as glm53-flash-ormandj-sglang-sm120-tp2-393k-c1-adaptive-mtp
  • Runtime identity: SGLang rc14 image sha256:0c0637959c3931829f05154087bbefd2c50003fb9b2010200ce0ec82f4d71a53
  • Local setup: two RTX PRO 6000 Blackwell Max-Q cards, exclusive TP=2 over PCIe/WSL2, ModelOpt W4A16/NVFP4, FP8 KV, adaptive EAGLE [3,5], C1
  • Recipe: managed 393K/C1 recipe
  • Measurement path: warm online direct qualification followed by managed, authenticated routed, and real-client acceptance
  • Headline result: 112.07 decode tok/s at 2,969 prompt tokens and 99.79 at 304,491 prompt tokens, each p50 of three C1 requests
  • Capability result: tools 20/20, coding 15/15, image/OCR 12/12, endurance 60/60, both thinking contracts, routed aliases, and real Pi/OpenClaw/Hermes passed
  • Important caveat: the 3,072 MiB/card standing reserve is waived for this model-only pair; post-qualification and post-client free VRAM were 2,101 and 2,543 MiB/card respectively; no video or C2 qualification
  • Decision: human-approved published current text/tools/image/OCR profile; exact 524K EXL3/DFlash2 service retained as rollback
  • Canonical evidence: https://fakoli.github.io/anvil-serving/findings/2026-09-02-glm53-sglang-sm120-393k-promotion/
  • Artifact set: artifact-manifest.json and README.md

X / short post

Recount immediately before posting.

GLM-5.3-Flash W4A16, 2x RTX PRO 6000 TP=2/C1: 112.1 tok/s at 4K, 99.8 at 304K; tools 20/20, image/OCR 12/12. 393K promoted. https://fakoli.github.io/anvil-serving/findings/2026-09-02-glm53-sglang-sm120-393k-promotion/

Reddit

Check the target community's current rules before posting.

GLM-5.3-Flash W4A16 on dual RTX PRO 6000: 393K/C1, 112 tok/s, clients pass

I qualified and promoted the pinned ormandj GLM-5.3-Flash W4A16/NVFP4
checkpoint on two RTX PRO 6000 Blackwell Max-Q cards using exclusive TP=2 over
PCIe/WSL2, SGLang rc14, FP8 KV, adaptive EAGLE [3,5], and concurrency one.

Headline results:

- 4K target / 2,969 actual prompt tokens: 112.07 decode tok/s
- 380K target / 304,491 actual prompt tokens: 99.79 decode tok/s
- tools 20/20, coding 15/15, image/OCR 12/12, endurance 60/60
- managed/routed gates and real Pi, OpenClaw, and Hermes acceptance passed

The 3 GiB-per-card campaign reserve was explicitly waived because this GPU
pair runs only model workloads. Free VRAM was 2,101 MiB/card after the direct
qualification and 2,543 MiB/card after client acceptance, with no OOM,
restart, crash, or shared-memory residue. This is a bounded local C1 result,
not a universal model ranking or a C2/video qualification.

Full methodology, failures, and evidence:
https://fakoli.github.io/anvil-serving/findings/2026-09-02-glm53-sglang-sm120-393k-promotion/

Screenshot alt text

Benchmark result card for pinned GLM-5.3-Flash W4A16/NVFP4 on two RTX PRO 6000 Blackwell Max-Q GPUs at TP=2, 393,216 context, and concurrency one. It reports 112.07 decode tokens per second at 2,969 prompt tokens, 99.79 at 304,491 prompt tokens, full tool/image/OCR/endurance passes, and real-client acceptance. A caveat states that the 3 GiB reserve was waived for model-only GPUs and video/C2 are not qualified.

Claim ledger

Public claim Conditions Evidence
112.07 decode tok/s 2,969 prompt tokens; C1; p50 of three promotion finding · safe393k-capacity-c1-4k-r3.json
99.79 decode tok/s 304,491 prompt tokens; C1; p50 of three promotion finding · safe393k-capacity-c1-380k-r3.json
Tools, coding, media, and endurance passed 20/20, 15/15, 12/12, and 60/60 promotion finding · safe393k-preflight-all-low.json · safe393k-quality-coding-agent-v2-disabled-r3-visible2048.json · safe393k-multimodal-image-c1.json · safe393k-endurance-c1-4k-r60.json
Reserve was waived without changing the physical measurement default 3,072 MiB/card; effective 0; 2,101 MiB/card after qualification VRAM policy section · vram-policy-reclassification.json