Sonnet 5.5 Against DeepSeek on Two Sparks: 31.9 Minutes vs 88.1
Same voxel-world prompt, one try each. The cloud world scored 9.0, the local one 8.5.
The cloud was faster and slightly better. The local build cost nothing per run.
@niklaslenz_ai gave Claude Sonnet 5.5 and DeepSeek-V4.1-Flash-1017 the same voxel-world prompt, one try each. Sonnet ran in the cloud in his Claude Code CLI and built Skyhaven. DeepSeek ran on two DGX Sparks with @MiaAI_lab’s two-Spark recipe, EXL3 2.9 bpw on TensorFold at commit 370daf0, and built AURELIA.
| Sonnet 5.5, cloud | DeepSeek-1017, 2 Sparks | |
|---|---|---|
| Done in | 31.9 min | 88.1 min |
| Tokens written | 194,375 | 159,465 |
| Browser check | 9.0/10 | 8.5/10 |
| Cost | About $4.60 at API list prices | No API bill |
The setups differ, he notes. The Sparks ran in a locked test container with a 262,144-token answer limit. Sonnet ran in his own CLI with internet, his files and about 71,000 tokens of context before the first answer.
His pick: Skyhaven is the better world, “but no by that much.” If it has to run at home, Mia’s recipe gives you AURELIA with no API bill. @MiaAI_lab says she actually likes the DeepSeek world better.
Sources
- @niklaslenz_ai on X · Oct 23, 7:30 AM PT
- @MiaAI_lab on X · Oct 23, 7:33 AM PT