VeizikNVIDIA, Apple Silicon and edge builds are available now.Release notes →
veiziknative inference engine
For business Sign inDownload
Hardware

GeForce RTX 3090

Veizik on GeForce RTX 3090 — NVIDIA · CUDA · sm_86.

Overview

What this device is.

GeForce RTX 3090Available
Vendor
NVIDIA
Backend
CUDA
Platform
Linux on x86-64
Generation
Ampere
Compute capability
sm_86
Availability
Available

On GeForce RTX 3090, 3 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_86 (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 familyEvidence on this deviceAvailability
FLUX.1
Image		Expected, untested Expected, untestedAvailable
Stable Diffusion 3.5
Image		Expected, untested Expected, untestedAvailable
Wan 2.1 · Wan 2.2
Video		Verified VerifiedAvailable
Qwen3.5
Language		Verified VerifiedAvailable
Qwen3.6
Language		Verified VerifiedAvailable

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).

ModelModeWorkloadtok/s
Qwen3.5-0.8Bbalanced20 / 128 tok967.5 tok/s · 2 runs · range 965.9–969.2
Qwen3.5-0.8Bbalanced20 / 1024 tok1,052 tok/s · single run
Qwen3.5-0.8Bfast20 / 128 tok972.0 tok/s · 2 runs · range 967.7–976.4
Qwen3.5-0.8Bquality20 / 128 tok880.8 tok/s · 2 runs · range 877.6–883.9
Qwen3.5-2Bbalanced20 / 128 tok586.5 tok/s · 2 runs · range 584.8–588.2
Qwen3.5-2Bbalanced20 / 1024 tok560.4 tok/s · single run
Qwen3.5-2Bfast20 / 128 tok585.1 tok/s · 2 runs · range 584.2–586.0
Qwen3.5-2Bquality20 / 128 tok542.7 tok/s · 2 runs · range 542.5–543.0
Qwen3.5-4Bbalanced20 / 128 tok376.8 tok/s · 2 runs · range 376.5–377.2
Qwen3.5-4Bbalanced20 / 1024 tok397.8 tok/s · single run
Qwen3.5-4Bfast20 / 128 tok376.8 tok/s · 2 runs · range 376.2–377.4
Qwen3.5-4Bquality20 / 128 tok342.1 tok/s · 2 runs · range 342.0–342.1
Qwen3.5-9Bbalanced20 / 128 tok190.2 tok/s · 2 runs · range 190.1–190.3
Qwen3.5-9Bbalanced20 / 1024 tok170.2 tok/s · single run
Qwen3.5-9Bfast20 / 128 tok190.4 tok/s · 2 runs · range 190.4–190.4
Qwen3.5-9Bquality20 / 128 tok172.8 tok/s · 2 runs · range 172.6–172.9
Qwen3.6-35B-A3Bbalanced22 / 128 tok194.3 tok/s · 3 runs · range 194.0–195.4
Qwen3.6-35B-A3Bbalanced22 / 1024 tok209.7 tok/s · single run
Qwen3.6-35B-A3Bfast22 / 128 tok172.7 tok/s · 2 runs · range 172.3–173.1
Qwen3.6-35B-A3Bsingle-token decoding22 / 128 tok197.9 tok/s · 3 runs · range 197.6–200.7

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.

ModelModeWorkloadms
Qwen3.5-0.8Bbalanced20 / 128 tok4.2 ms · 2 runs · range 4.2–4.2
Qwen3.5-0.8Bbalanced21 / 1024 tok4.4 ms · single run
Qwen3.5-0.8Bfast20 / 128 tok4.1 ms · 2 runs · range 4.0–4.2
Qwen3.5-0.8Bquality20 / 128 tok4.1 ms · 2 runs · range 4.1–4.1
Qwen3.5-2Bbalanced20 / 128 tok6.6 ms · 2 runs · range 6.6–6.7
Qwen3.5-2Bbalanced21 / 1024 tok7.1 ms · single run
Qwen3.5-2Bfast20 / 128 tok6.7 ms · 2 runs · range 6.7–6.7
Qwen3.5-2Bquality20 / 128 tok6.9 ms · 2 runs · range 6.8–7.0
Qwen3.5-4Bbalanced20 / 128 tok13.5 ms · 2 runs · range 13.4–13.6
Qwen3.5-4Bbalanced21 / 1024 tok13.9 ms · single run
Qwen3.5-4Bfast20 / 128 tok13.1 ms · 2 runs · range 13.0–13.2
Qwen3.5-4Bquality20 / 128 tok14.0 ms · 2 runs · range 13.7–14.3
Qwen3.5-9Bbalanced20 / 128 tok21.7 ms · 2 runs · range 21.7–21.7
Qwen3.5-9Bbalanced21 / 1024 tok22.4 ms · single run
Qwen3.5-9Bfast20 / 128 tok21.8 ms · 2 runs · range 21.7–21.8
Qwen3.5-9Bquality20 / 128 tok23.2 ms · 2 runs · range 23.2–23.3
Qwen3.6-35B-A3Bbalanced22 / 128 tok107.2 ms · 3 runs · range 106.9–109.0
Qwen3.6-35B-A3Bfast22 / 128 tok107.6 ms · 2 runs · range 107.4–107.8

Server throughput Measured here

Aggregate generated tokens per second through veizik serve.

ModelModeWorkloadtok/s
Qwen3.5-0.8B1 request at a time36 / 128 tok531.9 tok/s · 3 runs · range 528.2–532.1
Qwen3.5-0.8B4 concurrent requests36 / 128 tok1,778 tok/s · 3 runs · range 1,776–1,780
Qwen3.5-0.8B8 concurrent requests36 / 128 tok3,125 tok/s · 3 runs · range 3,125–3,130
Qwen3.5-9B1 request at a time34 / 128 tok101.4 tok/s · 3 runs · range 101.4–101.4
Qwen3.5-9B4 concurrent requests34 / 128 tok366.4 tok/s · 3 runs · range 366.4–366.9
Qwen3.5-9B8 concurrent requests34 / 128 tok663.0 tok/s · 3 runs · range 662.7–663.3
Qwen3.6-35B-A3B1 request at a time34 / 128 tok160.9 tok/s · 3 runs · range 159.1–160.9
Qwen3.6-35B-A3B4 concurrent requests34 / 128 tok222.4 tok/s · 3 runs · range 221.3–224.4
Qwen3.6-35B-A3B8 concurrent requests34 / 128 tok230.6 tok/s · 3 runs · range 229.4–230.8

Time to first token Measured here

From request to the first streamed token through veizik serve.

ModelModeWorkloadms
Qwen3.5-0.8B1 request at a time36 / 128 tok24.6 ms · 3 runs · range 24.5–26.6
Qwen3.5-0.8B4 concurrent requests36 / 128 tok62.9 ms · 3 runs · range 62.6–63.9
Qwen3.5-0.8B8 concurrent requests36 / 128 tok92.6 ms · 3 runs · range 92.5–92.8
Qwen3.5-9B1 request at a time34 / 128 tok55.6 ms · 3 runs · range 55.5–55.8
Qwen3.5-9B4 concurrent requests34 / 128 tok176.2 ms · 3 runs · range 133.1–176.4
Qwen3.5-9B8 concurrent requests34 / 128 tok308.8 ms · 3 runs · range 308.4–308.9
Qwen3.6-35B-A3B1 request at a time34 / 128 tok182.7 ms · 3 runs · range 182.1–185.8
Qwen3.6-35B-A3B4 concurrent requests34 / 128 tok625.5 ms · 3 runs · range 622.5–629.8
Qwen3.6-35B-A3B8 concurrent requests34 / 128 tok1,001.0 ms · 3 runs · range 998.5–1,003.3

GPU memory in use Measured here

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

ModelModeWorkloadGiB
Qwen3.5-0.8Bbalanced—2.27 GiB · single run
Qwen3.5-0.8Bfast—2.27 GiB · single run
Qwen3.5-0.8Bquality—2.42 GiB · single run
Qwen3.5-2Bbalanced—3.48 GiB · single run
Qwen3.5-2Bfast—3.48 GiB · single run
Qwen3.5-2Bquality—3.80 GiB · single run
Qwen3.5-4Bbalanced—5.49 GiB · single run
Qwen3.5-4Bfast—5.49 GiB · single run
Qwen3.5-4Bquality—6.27 GiB · single run
Qwen3.5-9Bbalanced—8.44 GiB · single run
Qwen3.5-9Bfast—8.44 GiB · single run
Qwen3.5-9Bquality—9.82 GiB · single run
Qwen3.6-35B-A3Bbalanced—21.44 GiB · single run
Qwen3.6-35B-A3Bfast—21.51 GiB · single run

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.

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 familyGeForce RTX 3090State
FLUX.1Expected, untestedExpected, untested
Stable Diffusion 3.5Expected, untestedExpected, untested
Wan 2.1 · Wan 2.2VerifiedVerified
Qwen3.5VerifiedVerified
Qwen3.6VerifiedVerified

Same architecture (sm_86), so a family verified on one carries to the others as Supported: NVIDIA GeForce RTX 3080 Ti, NVIDIA RTX A5000, NVIDIA A40.

Output

What the engine draws on this exact device.

Images rendered natively by veizik on this GeForce RTX 3090 — output, not a benchmark.

A serene alpine lake at dawn, mist over glassy water with pine forest and snow-capped peaks — AI-generated with SD3.5-large on veizik (NVIDIA RTX 3090).
SD3.5-largeNVIDIA GeForce RTX 3090
A futuristic city skyline at dusk with neon reflections on rain-wet streets — AI-generated with SD3.5-medium on veizik (NVIDIA RTX 3090).
SD3.5-mediumNVIDIA GeForce RTX 3090
A lone lighthouse on a rocky cliff in a dramatic storm with a shaft of light through moody clouds — AI-generated with SD3.5-large on veizik (NVIDIA RTX 3090).
SD3.5-largeNVIDIA GeForce RTX 3090
A cozy cabin in a golden birch forest with warm glowing windows and falling leaves — AI-generated with SD3.5-large on veizik (NVIDIA RTX 3090).
SD3.5-largeNVIDIA GeForce RTX 3090
A vibrant coral reef with tropical fish and sun rays through turquoise water — AI-generated with SD3.5-medium on veizik (NVIDIA RTX 3090).
SD3.5-mediumNVIDIA GeForce RTX 3090
A tranquil alpine lake at sunrise with mirror reflections of snow-capped peaks and a pine-forest shoreline — AI-generated with FLUX.1-schnell on veizik (NVIDIA RTX 3090).
FLUX.1-schnellNVIDIA RTX 3090

Shown as what the engine draws on a GeForce RTX 3090 — SD3.5 and FLUX.1 output, not a benchmark; no speed or memory figure is attached to these images. Each was rendered natively by the Veizik engine on this exact device. Powered by Stability AI.