FLUX.2 Klein 9B
@black-forest-labs/flux-2-klein-9b · NVIDIA B200 · August 2026
- compiled code size
- 16MB
- compiled code size
- artifact pull
- 1.6GB · 17 files
- artifact pull
- median cold-start time to first image
- 5.0s
- median cold-start time to first image
- median generation time · 1024×1024
- 426ms
- median generation time · 1024×1024
All stats above are computed from the raw result files linked at the bottom of this page. See methodology for how each metric is defined and measured.
Artifact size
The complete Muna-compiled runtime is 1.6GB across 17 files. Hover any slice to inspect the shared libraries that make up the artifact.
Muna-compiled FLUX.2 Klein 9B
Hover a slice to inspect a file.
Cold-start Time to First Image
Cold-start Time to First Image measures the full path to receiving the first 1024×1024 image. Muna results were captured on baremetal nodes, and each run dropped the page cache before every boot. SGLang results were measured on Modal, with and without GPU snapshotting. The GPU snapshot bar includes only verified restores and excludes snapshot bootstrap attempts.
Image generation time
Warm image-generation latency at batch size 1 across four square output resolutions. Each bar shows the median of 20 measured runs after three warmup runs.
Economic impact
A simple model of how rapid autoscaling changes monthly cost as demand rises and falls throughout the day. Even an always-on endpoint sees demand rise and fall throughout the day. SGLang pays for a warm GPU floor plus the startup time for autoscaled replicas. Muna can scale to zero, so its bill tracks useful inference and much shorter autoscale events.
$4,479
SGLang per month
$436
Muna per month
Illustrative assumptions · one B200 floor at $4,300/month · 24 autoscale events/day · measured cold starts: SGLang 151.7s, Muna 5.0s
Run configuration
- Model
- @black-forest-labs/flux-2-klein-9b
- Weights
- black-forest-labs/FLUX.2-klein-9B
- Hardware
- NVIDIA B200
- Run date
- August 2026
- Output resolution
- 1024×1024
- Diffusion steps
- 4
- Runtime
- nanosgl-cpp
- Platform
- runpod
- Storage
- local-nvme
- Page cache
- dropped before each cold boot
- Generation rounds
- 20 per resolution, 3 warmup, batch size 1
- Prompt
- A watercolor painting of a lighthouse at dawn on a rocky coast, with seabirds circling above crashing waves.
Raw data
Every chart on this page is rendered directly from these result files, including the raw per-run samples.
Generate with this compiled model.
Generate images with FLUX.2 Klein 9B on Muna, or deploy the same compiled model to your own compute.
$ muna deploy @black-forest-labs/flux-2-klein-9b --provider modal --gpu b200