Vade is a Rust inference engine for DGX Spark with zero third-party crates and exact repeatable output

Vade is an Apache-2.0 inference engine for NVIDIA’s DGX Spark desktop that is, in its README’s words, “written in Rust from the GPU kernels up”. It serves one model, Qwen3.8 Flash-Next, through an OpenAI-compatible API, and comes with its own terminal coding agent. Its pitch is exactness: “the same request, with the same settings, gives the same answer every time, and linking Sparks together does not change that.”

No dependencies, checked

The project’s design notes say the engine, CLI, harness and tools “use no third-party crates and no Python”, and that it does not use the CUDA toolkit, cuBLAS, CUTLASS or NCCL. Vade writes its own GPU kernels in Rust. We checked the Cargo.lock files for the engine, CLI and harness on main on 1 November: none lists a registry or git source, so every package is the project’s own.

What runs where

• Model: Qwen3.8 Flash-Next, text only, at pinned revisions. The FP4 build fits on one Spark; the FP8 build needs two or more.
• Hardware: DGX Spark (GB10) on Linux, with a driver for CUDA 13.0 or newer. The model download is 102.26 GB.
• Sparks: FP4 on 1, 2 or 4 Sparks, or a chain of up to 24 over TCP (measured only up to four). FP8 on 2 or 4.
• API: OpenAI protocol on port 18080. The README says Claude Code, Codex and other coding tools connect with a short setup.
• Agent: vade chat reads, searches and edits files. On the Spark itself it does not run shell commands, because its shell tool needs the macOS sandbox.

How exact is exact

Vade says greedy requests give the same token IDs on every run, alone or in a batch, with or without the prefix cache, on FP4 across 1, 2 and 4 Sparks. FP8 is newer: since 31 October it covers 2 and 4 Sparks, and the pre-merge check “matched 66 of 66 replies”. The README is open about gaps. FP8 samples by default, so you must set temperature and the penalties to 0 for exact replies, and six-Spark FP8 and some draft modes are not covered. You can test your own setup with vade verify.

Speed, as the author measured it

The README’s table, measured on main on 31 October on the author’s own Sparks, gives 42.5 tokens a second for one chat on one Spark with speculative drafting on, 63.0 on four Sparks, and 179.0 total for eight chats on one Spark. These are single-maintainer numbers; the raw records are in the repo. The first tagged release, v0.1.0, is dated 24 October, and the project is moving fast.

Vade joins a run of Rust-first GPU work we have covered: NVIDIA’s cuda-oxide 0.3, Respan’s Rust and CUDA Span-02 server and Resin.

Why this matters

Most serving stacks give slightly different tokens from run to run and across GPU counts, which makes regressions and evals hard to pin down. Vade shows a whole Rust path, from driver calls to kernels to server, built to make output repeatable, and it lets you check that claim with one command. If you own a Spark, it is a cheap test; just weigh that it serves one model and depends on one maintainer.