Hardware

NVIDIA GeForce RTX 4090

Veizik on NVIDIA GeForce RTX 4090 — NVIDIA · CUDA · sm_89.

Overview

What this device is.

NVIDIA GeForce RTX 4090Available

Vendor: NVIDIA
Backend: CUDA
Platform: Linux on x86-64
Generation: Ada Lovelace
Compute capability: sm_89
Availability: Available

On NVIDIA GeForce RTX 4090, 2 model families are verified end to end on a published build. Evidence and availability are separate: “verified” is a first-hand run on this exact device, not a download, and “available” never means verified on your hardware.

All devices · Platform requirements · Benchmarks & method

Engines

The engine that runs here.

Veizik native engineAvailable

Backend: CUDA
Architecture: sm_89 (one kernel set per architecture)
Model path: dep0 — pure native, no Python runtime
Families carried: FLUX.1, Stable Diffusion 3.5, Wan 2.1 · Wan 2.2, Qwen3.5, Qwen3.6

One engine runs every model family on this backend. Each GPU architecture is carried as its own native kernel set; a family is “verified” on a device only when it has been run there end to end.

Models

What runs on this device.

Pick a device, get a straight answer — which model families are verified, carried by architecture, or in preparation.

VerifiedVerifiedRun end to end on this exact device, on a published build.Architecture-supportedArchitecture-supportedVerified on another device of the same architecture; this device itself is untested.Expected, untestedExpected, untestedThe engine targets this architecture; nobody has run this model on it. No claim is made.In preparationIn preparationNo build is downloadable yet, and no figure is offered.Not offeredNot offeredThis family is not offered on this platform.UnknownUnknownNo information yet — the platform is announced with the release.

Model family	Evidence on this device	Availability
FLUX.1 Image	Expected, untested Expected, untested	Available
Stable Diffusion 3.5 Image	Expected, untested Expected, untested	Available
Wan 2.1 · Wan 2.2 Video	Expected, untested Expected, untested	Available
Qwen3.5 Language	Verified Verified	Available
Qwen3.6 Language	Verified Verified	Available

Stable Diffusion 3.5: Powered by Stability AI

Performance

Measured on this exact device.

Decode, tokens per second Measured here

Generated tokens per second, greedy, one request, through veizik run. Median of the runs shown (hover for the range).

Model	Mode	Workload	tok/s
Qwen3.5-0.8B	balanced	21 / 128 tok	1,318 tok/s · 2 runs · range 1,316–1,320
Qwen3.5-0.8B	balanced	21 / 1024 tok	1,439 tok/s · single run
Qwen3.5-0.8B	fast	21 / 128 tok	1,316 tok/s · 2 runs · range 1,315–1,317
Qwen3.5-0.8B	quality	21 / 128 tok	1,156 tok/s · 2 runs · range 1,155–1,156
Qwen3.5-2B	balanced	21 / 128 tok	886.9 tok/s · 2 runs · range 879.2–894.7
Qwen3.5-2B	balanced	21 / 1024 tok	868.8 tok/s · single run
Qwen3.5-2B	fast	21 / 128 tok	895.1 tok/s · 2 runs · range 895.0–895.2
Qwen3.5-2B	quality	21 / 128 tok	740.9 tok/s · 2 runs · range 740.9–741.0
Qwen3.5-4B	balanced	21 / 128 tok	531.8 tok/s · 2 runs · range 528.6–535.0
Qwen3.5-4B	balanced	21 / 1024 tok	568.2 tok/s · single run
Qwen3.5-4B	fast	21 / 128 tok	535.4 tok/s · 2 runs · range 535.1–535.7
Qwen3.5-4B	quality	21 / 128 tok	441.8 tok/s · 2 runs · range 441.8–441.8
Qwen3.5-9B	balanced	21 / 128 tok	347.0 tok/s · 2 runs · range 346.7–347.3
Qwen3.5-9B	balanced	21 / 1024 tok	303.4 tok/s · single run
Qwen3.5-9B	fast	21 / 128 tok	346.6 tok/s · 2 runs · range 346.1–347.2
Qwen3.5-9B	quality	21 / 128 tok	272.8 tok/s · 2 runs · range 272.8–272.8

Prompt pass Measured here

Time for the prompt pass of the short fixed prompt, including the first generated token, as the engine reports it. Not a prefill throughput.

Model	Mode	Workload	ms
Qwen3.5-0.8B	balanced	21 / 128 tok	3.0 ms · 2 runs · range 3.0–3.0
Qwen3.5-0.8B	balanced	21 / 1024 tok	3.0 ms · single run
Qwen3.5-0.8B	fast	21 / 128 tok	3.0 ms · 2 runs · range 3.0–3.0
Qwen3.5-0.8B	quality	21 / 128 tok	3.1 ms · 2 runs · range 3.1–3.1
Qwen3.5-2B	balanced	21 / 128 tok	4.5 ms · 2 runs · range 4.4–4.6
Qwen3.5-2B	balanced	21 / 1024 tok	4.4 ms · single run
Qwen3.5-2B	fast	21 / 128 tok	4.4 ms · 2 runs · range 4.4–4.4
Qwen3.5-2B	quality	21 / 128 tok	4.7 ms · 2 runs · range 4.7–4.7
Qwen3.5-4B	balanced	21 / 128 tok	8.7 ms · 2 runs · range 8.6–8.9
Qwen3.5-4B	balanced	21 / 1024 tok	8.6 ms · single run
Qwen3.5-4B	fast	21 / 128 tok	8.7 ms · 2 runs · range 8.6–8.7
Qwen3.5-4B	quality	21 / 128 tok	10.5 ms · 2 runs · range 10.5–10.5
Qwen3.5-9B	balanced	21 / 128 tok	16.8 ms · 2 runs · range 16.7–16.9
Qwen3.5-9B	balanced	21 / 1024 tok	16.6 ms · single run
Qwen3.5-9B	fast	21 / 128 tok	16.5 ms · 2 runs · range 16.5–16.5
Qwen3.5-9B	quality	21 / 128 tok	21.6 ms · 2 runs · range 21.6–21.6

GPU memory in use Measured here

Once the weights and the decode graph are loaded — not the transient peak.

Model	Mode	Workload	GiB
Qwen3.5-0.8B	balanced	—	2.40 GiB · 2 runs
Qwen3.5-0.8B	fast	—	2.40 GiB · 2 runs
Qwen3.5-0.8B	quality	—	2.55 GiB · 2 runs
Qwen3.5-2B	balanced	—	3.61 GiB · 2 runs
Qwen3.5-2B	fast	—	3.61 GiB · 2 runs
Qwen3.5-2B	quality	—	3.93 GiB · 2 runs
Qwen3.5-4B	balanced	—	5.62 GiB · 2 runs
Qwen3.5-4B	fast	—	5.62 GiB · 2 runs
Qwen3.5-4B	quality	—	6.40 GiB · 2 runs
Qwen3.5-9B	balanced	—	8.57 GiB · 2 runs
Qwen3.5-9B	fast	—	8.57 GiB · 2 runs
Qwen3.5-9B	quality	—	9.95 GiB · 2 runs

Every figure is veizik’s own measurement on this exact device, bound to the published build it was taken on. See Reproduce our benchmarks.

Compatibility

Architecture and its siblings.

Model family	NVIDIA GeForce RTX 4090	State
FLUX.1	Expected, untested	Expected, untested
Stable Diffusion 3.5	Expected, untested	Expected, untested
Wan 2.1 · Wan 2.2	Expected, untested	Expected, untested
Qwen3.5	Verified	Verified
Qwen3.6	Verified	Verified

No sibling device of this architecture is in the matrix.