GLM-5.3-Flash on Two Sparks: 49.9 Tok/s and a Sky Kingdom in 71.1 Minutes
@niklaslenz_ai measured @MiaAI_lab’s recipe on the TensorFold Python engine; her Zig branch was slower.
@niklaslenz_ai ran GLM-5.3-Flash on two DGX Sparks with @MiaAI_lab’s recipe at commit 3fd60dfc on the TensorFold Python engine. The recipe uses an EXL3 4 bpw quant, one model split across both Sparks (TP2), a DFlash2 draft model and vision on.
| Measure | Result |
|---|---|
| Decode, code | 51.2 tok/s |
| Decode, prose | 48.6 tok/s |
| Draft acceptance | 61.7% |
| Reading a 32K-token prompt | 1,944 tok/s |
| First token after a 32K prompt | 16.9 s |
| Eight requests at once, total | 99.2 tok/s |
| MMLU-Pro | 64.8% ± 4.7 points |
| GPQA with thinking | 73.3% ± 15.1 points |
As an agent in OpenCode 1.19.3, the setup built Aetherfall, a voxel sky kingdom of floating isles with a sky whale, in 71.1 min. It wrote 109,305 tokens and scored 9/10 on his browser check. He notes parts of the world are cut off hard at the edges, most likely a bug in the generated code.
He ran the same task first on Mia’s Zig-engine branch, which her docs call work in progress. It decoded code at 26.6 tok/s and took 98.8 min.
@bertholomusai calls it a “GOAT recipe.” GLM 5.3 Flash has been his daily driver on two Sparks for almost a month, he says, and he now has it running on four.
Sources
- @niklaslenz_ai on X · Oct 24, 9:00 AM PT
- @niklaslenz_ai on X · Oct 24, 9:30 AM PT
- @bertholomusai on X · Oct 25, 12:16 AM PT