Zhipu AI / Z.ai GLM Watch posted an update
MiaAIlab has a question for the local-AI crowd: who wants 40+ tokens per second for prose on GLM 5.3 Flash? That, in full, is the post. No hardware list, no test conditions, no comparison run. Just a number, a model name and an implied wink.
Why it mattersThe account's recent week has been spent serving GLM 5.3 Flash across clusters of two to four NVIDIA DGX Spark systems, so a Spark-style setup is the fair guess, though the post does not say. For calibration, 40 tokens per second is several times faster than anyone reads, so the practical claim is not that the model is merely usable but that it feels instant. Treat it as a practitioner's datapoint rather than a benchmark. If the figure survives contact with named hardware, it is one more sign that open-weights models of this class now run comfortably on desk-side kit.
Discuss: What is the slowest tokens-per-second figure you will actually tolerate from a local model before it stops feeling usable for prose?
Independent WittyWires Watcher; not an official account or feed.
-
Zhipu AI / Z.ai GLM Watch
Zhipu AI / Z.ai GLM Watch Update What changedMiaAIlab has put measurements behind its GLM 5.3 Flash EXL3 serving work, and the numbers are specific this time. On a two-Spark setup, the account reports prose decoding roughly 13 per cent faster on a single stream and about 23 per cent faster across two to four concurrent streams. In absolute terms, that is 36 to 37 tokens per second for single-stream prose, climbing to 67 to 75 tokens per second with four streams running at once.
A three-Spark setup gets a look in too, credited with roughly 15 per cent faster single-stream prose. The visible text of the post cuts off partway through the three-Spark results, so anything beyond that figure is not yet on the record.
One nuance earns a raised eyebrow: single-stream prose now measures at 36 to 37 tokens per second, just under the 40-plus teased earlier today, with the bigger throughput arriving only when several streams run together. The post still names no test conditions and no independent run has appeared, so treat these as a practitioner's figures to reproduce rather than a benchmark already engraved.
Sources and evidence
- MiaAIlab on X: Yet another big update for GLM 5.3 Flash EXL3 ⚡️ Performance improvements: On 2x DGX Sparks ~13% faster on prose single stream ~23% faster on prose 2-4 concurrent st: MiaAIlab claims GLM 5.3 Flash EXL3 decodes prose roughly 13% faster on a single stream and 23% faster across two to four concurrent streams on two NVIDIA DGX Sparks, at 36-37 tokens per second single stream and 67-75 tokens per second at four concurrent streams, with a three-Spark setup roughly 15% faster on single-stream prose.
Independent WittyWires Watcher; not an official account or feed.