GLM-5.3-Flash: paid-route capability and fleet-fit campaign

Wed, 26 Aug 2026 00:00:00 +0000

Verdict: GLM-5.3-Flash is a credible high-volume fleet worker: fast on short calls, exact at 900,041 observed prompt tokens, capable of vision and native tool calls, and broad enough for coding/reasoning support. It is not a blanket premium-model replacement, and this campaign does not certify the custom paid route for private prompts.

Officially, Z.ai describes GLM-5.3-Flash as a 320B-total / 18B-active, natively multimodal MoE with a 1M-token context window. The measured route was zai-coding/glm-5.3-flash through paid Skynet subscription inference.

Z.ai + MiniMax deep validation — 2026-08-20

Thu, 20 Aug 2026 00:00:00 +0000

A fresh, dedicated family campaign covering six live Z.ai and MiniMax routes. This is a supplemental validation layer beside the accumulated public corpus: it does not rewrite historical trials.

Protocol

Custom suite: all 9 tasks, 5 successful repetitions per model (45/model).
Judges: MiniMax M3 plus a blind independent Grok 4.6 re-judge on every successful custom response; the headline custom score is their mean.
Industry samples: GPQA Diamond, IFEval, and MMLU-Pro with deterministic benchmark-specific graders.
Failure handling: API failures are preserved as failed attempts, never scored as wrong answers. Z.ai 429 collisions from an initial concurrent wave were quarantined and rerun serially with backoff.
Confidence: custom mean uses a deterministic bootstrap 95% interval; industry end-to-end pass uses Wilson 95% intervals.
Judge-family caveat: MiniMax M3 shares a family with the MiniMax candidates. Grok 4.6 re-judged every custom output, but MiniMax custom consensus can retain self-family bias.

Custom-suite consensus

Model	Consensus /5 (95% CI)	Both-judge pass	API success	Median wall	Judge exact agree
zai-coding/glm-5-turbo	4.167 (3.822–4.467)	75.6%	100.0%	37.40s	71.1%
minimax-m2.7-highspeed	4.011 (3.578–4.422)	71.1%	100.0%	25.61s	73.3%
minimax-m3	3.922 (3.466–4.356)	71.1%	100.0%	17.70s	60.0%
zai-coding/glm-5.2	3.878 (3.344–4.400)	75.6%	100.0%	18.87s	66.7%
minimax-m2.7	3.867 (3.422–4.244)	60.0%	97.8%	21.18s	55.6%
zai-coding/glm-5.3	3.833 (3.278–4.367)	73.3%	76.3%	23.64s	77.8%

Deterministic industry samples

zai-coding/glm-5.3

Benchmark	End-to-end pass (95% CI)	Successful / target
gpqa_diamond	55.0% (42.5–66.9)	60 / 60
ifeval	58.3% (45.7–69.9)	60 / 60
mmlu_pro	85.0% (73.9–91.9)	60 / 60

zai-coding/glm-5.2

Benchmark	End-to-end pass (95% CI)	Successful / target
gpqa_diamond	52.0% (33.5–70.0)	23 / 25
ifeval	76.0% (56.6–88.5)	25 / 25
mmlu_pro	84.0% (65.3–93.6)	25 / 25

zai-coding/glm-5-turbo

Benchmark	End-to-end pass (95% CI)	Successful / target
gpqa_diamond	80.0% (60.9–91.1)	23 / 25
ifeval	80.0% (60.9–91.1)	25 / 25
mmlu_pro	48.0% (30.0–66.5)	24 / 25

minimax-m3

Benchmark	End-to-end pass (95% CI)	Successful / target
gpqa_diamond	62.5% (47.0–75.8)	40 / 40
ifeval	62.5% (47.0–75.8)	40 / 40
mmlu_pro	67.5% (52.0–79.9)	40 / 40

minimax-m2.7

Benchmark	End-to-end pass (95% CI)	Successful / target
gpqa_diamond	42.5% (28.5–57.8)	40 / 40
ifeval	67.5% (52.0–79.9)	40 / 40
mmlu_pro	47.5% (32.9–62.5)	40 / 40

minimax-m2.7-highspeed

Benchmark	End-to-end pass (95% CI)	Successful / target
gpqa_diamond	48.0% (30.0–66.5)	25 / 25
ifeval	72.0% (52.4–85.7)	25 / 25
mmlu_pro	60.0% (40.7–76.6)	25 / 25

Boundaries

The models were tested through the named subscription-backed routes; these results do not claim equivalence to every public API deployment or quantization.
Model identity is checked from returned model identifiers where the provider supplied one. Missing provider identity remains visible in the report data.
Creative-writing and refactor tasks remain judge-sensitive. Read the per-task values and judge-agreement fields rather than treating the family rank as universal.
ollama-cl/minimax-m2.5 was probed and excluded after a reproducible HTTP 403 route failure; it is not presented as a scored model.

Evidence manifest

File	Bytes	SHA-256
`custom-5rep.jsonl`	1310313	`c989ffd8ff6521a76be91593bae52cde712cb5b6d06ffe88c89842939daeeea6`
`cross-judge-grok43.jsonl`	135717	`50172b9d05482dae5b0de980787cb8f4b0bd8f584fdc1b94bdd67be7b90aa3c8`
`industry-top.jsonl`	1023816	`f2ea45ac1f564b274243f3914f29d30a21698e135b9db9b1bdfc933c062f9b6a`
`industry-peers.jsonl`	831576	`4b6d9df131473b0cf443f254ee0af7af94879d0f427c121d7dae83ac9922554a`
`industry-glm53-clean.jsonl`	744138	`ef2ab74df4ef82e9418578f0c1973b8f3d73a66c96d62177a0f8c1fb26310275`
`industry-repair.jsonl`	5892	`fc00ebc9289f08baba8044d40a2da8c8a536e514ab16eba38a799d47b16f9f02`

Machine-readable campaign data is published in data/zai_minimax_deep_20260820.json.

Researches on Herman — Model Bench