GLM-5.3 / 11 SEPTEMBER 2026
Less waiting.
More useful work.
4.02 tok/s versus 3.46 tok/s on the same four-drive engine, across three matched 512-token pairs.
What was tested
Full GLM-5.3 (744B MoE), using the UD-Q4_K_XL base with routed experts requantized to uniform Q4_K. Model size: 434,170,891,168 bytes. M5 Max, 128 GB; internal SSD plus three NVMe enclosures. The same ds4 base (6289c51) and patched binary were used on both sides.
The candidate enables poll-gated pre-commit and a dense eviction scan. The control disables those two changes. Both retain split reads, lookahead prefetch, the same automatic expert cache and a 10:5:5:4 replica split with 48 reader threads. This isolates two scheduling/cache changes; it is not a comparison against unmodified upstream.
Measurement
Raw prompt: “The three main financial statements are” (24 tokens including BOS), greedy decoding, no thinking. Fresh engine process and empty application expert cache for each arm. Decode rate measures tokens after the first response divided by time from first to last response; it excludes prefill and startup. Raw metadata records the dashboard as open across all twelve arms. Existing swap was allowed; every arm recorded zero swap growth.
| Arm | Decode tok/s | First response, s | Output MD5 prefix |
|---|---|---|---|
| qual128-cand-1 | 4.2130 | 4.51 | da46fa88983a… |
| qual128-ctrl-1 | 3.6602 | 4.56 | da46fa88983a… |
| qual128-cand-2 | 4.2215 | 4.44 | da46fa88983a… |
| qual128-ctrl-2 | 3.6581 | 4.51 | da46fa88983a… |
| qual128-cand-3 | 4.2134 | 4.56 | da46fa88983a… |
| qual128-ctrl-3 | 3.6612 | 4.70 | da46fa88983a… |
| qual512-cand-1 | 4.0161 | 4.61 | 6cc7187b3673… |
| qual512-ctrl-1 | 3.4610 | 4.63 | 6cc7187b3673… |
| qual512-cand-2 | 3.9982 | 4.59 | 6cc7187b3673… |
| qual512-ctrl-2 | 3.4688 | 4.61 | 6cc7187b3673… |
| qual512-cand-3 | 4.0578 | 4.53 | 6cc7187b3673… |
| qual512-ctrl-3 | 3.4416 | 4.63 | 6cc7187b3673… |
Results and limits
At 128 generated tokens, candidate median 4.2134 versus control 3.6602 tok/s (+15.1%). At 512 tokens, 4.0161 versus 3.4610 (+16.0%). Output hashes match across every six-arm group at a given length.
These are dated campaign results. A candidate-only recheck on 12 September measured 3.8609 tok/s median at 512 tokens; it had no new matched control. Performance varies with run conditions, and no causal explanation for that difference is established here.
Output equality between these scheduling variants does not establish requantization quality against the original quantization or an FP8 reference. That comparison remains unqualified. This GLM metric, prompt and baseline differ from the DeepSeek V4.1 campaign.
Download sanitized measurements ↓
Data transcribed from the local qualification logs and output-hash files. The JSON includes source-log hashes, exact observations and conditions; it excludes machine paths, personal details and private notes. No claim is made that the public DeepSeek fork reproduces this separate GLM configuration.