27 OCTOBER 2026 / DEVELOPER BENCHMARKS

Here are the numbers.

Elapsed seconds. Lower is better. Primary shapes requested 2 allocated vCPUs and 4 GiB RAM.

Workload	Luisa	Boat	exe.dev	E2B	Freestyle	Daytona
Rust: fd clean release	48.02	61.61	71.09	106.24	106.09	81.46
Rust: ripgrep clean release	21.88	24.87	30.32	42.12	40.19	34.29
Rust: edit, compile + 135 tests	1.49	2.59	2.30	3.19	3.89	2.84
Excalidraw TypeScript full check	6.57	12.90	12.79	17.81	22.93	13.91
Excalidraw Vite production build	15.87	38.79 †	27.26	39.37 †	43.84	37.29
Java: Commons Lang clean compile	6.64	13.83	13.16	20.04	20.71	28.46
Java: edit, compile + 12 tests	7.25	14.35	14.34	20.87	22.60	34.44
Linux ARM64 defconfig Image	686.31	1036.27	1524.46	1941.05	2545.43	N/A

† Boat’s Vite result has one complete instance; E2B’s has two. These partial results are excluded from the landing-page comparisons. Daytona could not supply the frozen kernel compiler; no substitute score is included.

How we measured

Three fresh instances per provider. Developer workloads used three repetitions per instance; kernel workloads used one full build per instance. The reported result is the median of the per-instance medians.
Luisa used an isolated development runtime on one Apple M4 host, running native ARM64 Linux. Other providers used x86-64; Daytona was explicitly a container. Three Luisa VMs do not mean three independent physical hosts.
Source revisions and toolchains were pinned. Setup and downloads are excluded. Clean builds started with build outputs removed, while ordinary OS caches were retained. Incremental compilation was used only for the labelled Rust edit case.
The Linux workload built Linux 6.18.59 to the same ARM64 defconfig Image target with the same LLVM upstream revision and semantic config.

What the results do—and don’t—say

Requests for CPU count and physical RAM match. Architecture, memory backing, disk allowances and base-image services differ. Luisa retained 16 GiB swap, Boat 2 GiB, and other observed guests zero.
These measurements do not isolate CPU or NVMe as the cause of the differences. They do not qualify Luisa’s production service or public gateway, establish a general fastest-provider claim, or guarantee future performance.
The multiplier on each landing-page card is the runner-up’s time ÷ Luisa’s time, to one decimal. All providers in that workload remain visible. Kernel values are shown in minutes there; the table above keeps seconds.