How Seinari benchmarked on up to 16 RTX 5090s at once on Lium

What they rented

• GPU: RTX 5090, 32 GB, one per pod.
• Scale: 120 pods, up to 16 running at once, 165 GPU-hours in total.
• Tier: 81 Spot pods and 39 Secure pods.
• Price paid: $0.72 per GPU-hour on average on Spot, $0.79 on Secure.
• Setup: self-serve from the console. No quota request, sales call or contract.

Price per GPU-hour

Why RTX 5090 on Lium

• Right card: 32 GB fits a 35B mixture-of-experts model at 4-bit and the FLUX image models on one GPU.
• Many at once: 260 RTX 5090s were free on Lium at the read time above, so 16 parallel runs needed no waiting.
• Short runs cost little: per-second billing, so a five-minute benchmark costs five minutes.

In Seinari's words

Seinari builds a superoptimizer for AI: finding the fastest program for a model on a given chip. With support from Lium's research credits, the team measured its current inference engines on Lium RTX 5090 pods:

• Qwen3.6-35B-A3B Q4: 1,042 decode tokens/s for a single stream.
• FLUX.2 Klein: 79 ms at 512 × 512 in Fast mode; 407 ms at 1024 × 1024 in Quality mode.
• FLUX 3 Klein: 111 ms at 512 × 512 in Fast mode; 1.283 seconds at 1024 × 1024 in Quality mode.

Qwen's figure is the geometric mean of eight case medians, with three 512-output-token runs per case; prompt prefilling is excluded. Image figures are medians of eight warm local requests, including text preparation and full image decoding. FLUX.2 Fast uses three steps; Quality uses four. FLUX 3 Fast uses four steps; Quality uses sixteen.

Try the image generator and local model runner at seinari.ai.

Run it yourself on Lium

curl -fsSL https://lium.io/install.sh | bash
lium init

lium up --gpu RTX5090 --name bench --ttl 2h -y

lium rm bench -y