Can I Run Cosmos 3 on RTX 5090 Desktop (32 GB VRAM, 64 GB RAM)?

Autor: Jakub Rusinowski · Ostatnia aktualizacja: 21 lipca 2026

Yes — comfortably

Yes, comfortably — Cosmos 3 Nano at Q8_0 needs about 19.2 GB of the 32 GB usable on RTX 5090 Desktop (32 GB VRAM, 64 GB RAM), leaving ~12.8 GB spare and running at ~109.8 tok/s (estimated), with room for about 16,384 tokens of context.

Confidence: medium · Recommended quantization: Q8_0 · Estimated speed: ~109.8 tok/s

Buy This HardwareAMD Radeon RX 7900 XT 20GB — 20 GB VRAM · 315 W board power Deploy in the Cloud NowRTX 4090 on RunPod — from $0.34/hr · rate checked 2026-07

or compare on Vast.ai from $0.35/hr (typical low · varies)

As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.

RTX 5090 Desktop (32 GB VRAM, 64 GB RAM) — what it gives a model

Usable memory for models	32 GB
Memory bandwidth	1792 GB/s
Form factor	Desktop
Operating system	Windows or Linux
Memory upgradeable	Yes

Cosmos 3 on RTX 5090 Desktop (32 GB VRAM, 64 GB RAM): memory by quantization

Quant	Memory needed	Fits 32 GB?	Max context	Est. speed	Download
F16	34.2 GB	✗ No	—	—	32 GB
Q8_0	19.2 GB	✓ Yes	16K	~109.8 tok/s	17 GB
Q6_K	15.4 GB	✓ Yes	16K	~130.5 tok/s	13.1 GB
Q5_K_M	13.6 GB	✓ Yes	16K	~142.8 tok/s	11.3 GB
Q4_K_M	11.9 GB	✓ Yes	16K	~156.8 tok/s	9.7 GB
Q3_K_M	9.1 GB	✓ Yes	16K	~188 tok/s	6.8 GB
Q2_K	7.5 GB	✓ Yes	16K	~211 tok/s	5.3 GB

Which Cosmos 3 sizes fit

Variant	Needs at Q4_K_M	Fits?	Est. speed
Cosmos 3 Super	41.8 GB	✗ Too large	—
Cosmos 3 Nano	11.9 GB	✓ Fits	~156.8 tok/s
Cosmos 3 Edge	4.1 GB	✓ Fits	~287.2 tok/s

What to watch out for

1 larger variant of Cosmos 3 does not fit and would need CPU offload or different hardware.
This model does not publish its full attention configuration, so the KV-cache share of these figures is inferred from its parameter count rather than computed exactly.

RTX 5090 desktop limitations

575 W board power — budget for a 1000 W+ PSU and the heat it puts into the room.
Models larger than 32 GB must offload to system RAM, which costs roughly an order of magnitude in speed.

Recommended setup

Ollama or llama.cpp (CUDA/ROCm) — vLLM if you need concurrent serving

How these numbers are calculated

Memory need = quantized weights + KV cache at 8,192 tokens (f16) + 0.8 GB runtime overhead.
32 GB of VRAM on the NVIDIA GeForce RTX 5090 at 1792 GB/s.
64 GB of system RAM available for CPU offload when a model exceeds VRAM.
Throughput is a memory-bandwidth roofline estimate, not a measurement. It is labelled "estimated" everywhere it appears.
KV cache is inferred from the parameter count because this model does not publish its full attention configuration.

FAQ