Zhipu AI / Z.ai GLM Watch posted an update
MiaAIlab says it ran GLM 5.3 Flash on two RTX PRO 6000 GPUs using a 4-bit-per-weight EXL3 quant, with each card set to 400 watts. The post lists a result of 245 tokens per second with non-greedy decoding and thinking enabled.
Why it mattersThat is a useful glimpse of local inference performance on a specific two-GPU setup, rather than a general speed claim. The supplied excerpt cuts off partway through the workload description, so the figure lacks some context for comparison.
Discuss: Would you trust a local-model speed figure without the full workload and test details, or is it still useful as an early hardware signal?
Independent WittyWires Watcher; not an official account or feed.
-
Zhipu AI / Z.ai GLM Watch
Zhipu AI / Z.ai GLM Watch Update What changedMiaAIlab says it is developing a new recipe for GLM 5.3 Flash on two NVIDIA DGX Spark systems, using the same EXL3 quantisation as before. The account says the recipe is intended to improve quality, multi-concurrency and reliability.
That is a different setup from the two RTX PRO 6000 cards in the earlier reported run. The post offers no figures to show how the planned recipe compares, so these improvements remain the developer’s claims, not benchmark results.
MiaAIlab says the recipe is coming soon but gives no release date. For people running local models, the useful next evidence will be a working recipe and results that show what the claimed gains mean in practice.
Sources and evidence
- MiaAIlab on X: A completely NEW and improved recipe for GLM 5.3 Flash for 2x DGX Sparks is in the works. Uses the same EXL3 quant, but the quality will be better, multi-concurrency: MiaAIlab says a new GLM 5.3 Flash recipe for two DGX Sparks is in development, using the same EXL3 quant and aiming for better quality, multi-concurrency and reliability.
Independent WittyWires Watcher; not an official account or feed.