Skip to content

Publication summary: GLM-5.3-Flash and bounded dual-PRO qualification

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

Canonical facts

  • Model identity: brandonmusic/GLM-5.3-Flash-tr3-4bpw@5ab363a8dcf6405955fd5f99671e01a1c9fb124b; preferred served name glm53-flash-tr3-4bpw-tp2-262k-fixed-mtp5-vision
  • Runtime identity: vLLM 0.1.dev20051+g487ecf187; pinned Purtell image digest sha256:da5cec95778bf6996660b52e28a6e51737fec69cfc3d508bf298c8a89f273ac5
  • Local setup: 2x RTX PRO 6000 Blackwell Max-Q 96 GB, TP=2 over PCIe, Docker Desktop/WSL2, vision fixed K5, 262K, c1 performance
  • Recipe: managed vision fixed-K5 262K recipe
  • Measurement path: warm direct OpenAI-compatible API; no router or client path
  • Headline result: vision fixed K5 decoded at 72.8 tok/s at 4K and 55.7 tok/s at 128K, each from three c1 low-reasoning repetitions
  • Capability result: image understanding, verbatim OCR, tools 20/20, bounded high-reasoning coding 15/15, and exact retrieval at a 250K target / 206,296 actual prompt tokens passed; the no-spec companion recovered an exact needle at 495,045 prompt tokens
  • Important caveat: video is disabled, only one-image behavior is proven, adaptive K1-K5 plus ReplaySSM corrupted repeated tools, and 0xSero 3.0-bpw has no complete serve path and failed its publisher's held-out quality gate
  • Decision: challenger, no-promotion; no route, client, or promotion state changed
  • Canonical evidence: https://github.com/fakoli/anvil-serving/blob/main/docs/findings/2026-08-29-glm53-cardillo-purtell-qualification.md

X / short post

Dual-PRO GLM-5.3-Flash: vision/OCR, tools 20/20, 262K context, 72.8 tok/s. 3-bpw watch-only. No promotion. https://github.com/fakoli/anvil-serving/blob/main/docs/findings/2026-08-29-glm53-cardillo-purtell-qualification.md

Reddit

Dual RTX PRO 6000 GLM-5.3-Flash 4 bpw: vision/OCR and 262K serving qualified
I reproduced the Cardillo/Purtell GLM-5.3-Flash recipe on two RTX PRO
6000 Blackwell Max-Q cards under WSL2. Vision fixed K5 passed semantic image
understanding, verbatim OCR, 20/20 tools, exact retrieval at a 250K target /
206,296 actual prompt tokens,
and a bounded 15/15 high-reasoning coding suite. It reached 72.8 tok/s at 4K
and 55.7 at 128K. Video remains disabled. The no-spec 524K profile recovered
an exact needle at 495,045 prompt tokens and retained 1,603,111 reported KV
tokens. Adaptive MTP corrupted repeated tool output. The 0xSero 3.0-bpw lead
was not downloaded because its own release ledger lacks a complete server and
records a failed held-out quality gate. GLM remains unpromoted pending hands-on
and routed client acceptance. Full immutable identities, recipes, failures,
raw artifacts, and runtime-state proof are in the canonical finding:
https://github.com/fakoli/anvil-serving/blob/main/docs/findings/2026-08-29-glm53-cardillo-purtell-qualification.md

What matched or differed on your hardware?

Screenshot alt text

Benchmark result card for GLM-5.3-Flash TR3/EXL3 4 bpw on two 96 GB RTX PRO 6000 Blackwell Max-Q GPUs. Vision fixed five-token MTP passed image, OCR, tools, bounded coding, and a 250K-target / 206,296-actual retrieval while reaching about 73 output tokens per second at 4K. A no-speculation profile passed exact retrieval near 495K prompt tokens with three full windows of reported KV capacity. Adaptive MTP is marked rejected for tool corruption, 0xSero 3.0-bpw is watch-only, and the overall model remains no-promotion.

Claim ledger

Public claim Conditions Evidence
semantic image and verbatim OCR pass exact vision recipe; one image; video disabled matched local results; vision-fixed-mtp5-preflight-all-low.json
vision tools 20/20 and coding 15/15 direct API; high control recorded for coding but not independently verified matched local results; vision-fixed-mtp5-preflight-all-low.json and vision-fixed-mtp5-coding-agent-v2-high-r3.json
250K-target / 206,296-actual exact retrieval one calibrated request maximum context; vision-fixed-mtp5-needle-250k-low.json
vision fixed K5 72.8/55.7 tok/s c1 low reasoning, three repetitions matched local results; vision-fixed-mtp5-capacity-4k-c1-low-r3.json and vision-fixed-mtp5-capacity-128k-c1-low-r3.json
fixed K5 tools 20/20 exact model/image, direct API matched local results; fixed-mtp5-tools-low-r1.json and fixed-mtp5-524k-tools-high-r1.json
high-reasoning coding 15/15 bounded five-item diagnostic; request control recorded but not independently verified matched local results; fixed-mtp5-524k-coding-agent-v2-high-r3.json
495,045-token exact retrieval one calibrated request per profile maximum context; nospec-524k-needle-500k-actual-low.json and fixed-mtp5-524k-needle-500k-actual-low.json
497,976-token valid tool call one calibrated request per profile maximum context; nospec-524k-long-tool-500k-actual-low.json and fixed-mtp5-524k-long-tool-500k-actual-low.json
fixed K5 69.8/61.9 tok/s c1 low reasoning, three repetitions matched local results; fixed-mtp5-capacity-4k-c1-low-r3.json and fixed-mtp5-capacity-128k-c1-low-r3.json
adaptive MTP rejected repeated tool corruption and 13/15 coding attempts matched local results; preflight-tools-repeat01.json and adaptive-coding-agent-v2-low-r3.json