Can the RTX 5090 Run Qwen 2.5?
Yes — and very well. The RTX 5090’s 32 GB VRAM is the sweet spot for Qwen 2.5 14B at FP16 with a long context window. Larger Qwen variants require quantisation or a bigger card.
Yes — tightThe RTX 5090 (32 GB) just fits Qwen 2.5 (14B class) at FP16 (28 GB). For comfortable production use, drop to FP8 / INT8 quantisation.
Detailed Breakdown
The RTX 5090 with 32 GB of GDDR7 is one of the best single GPUs we host for the Qwen 2.5 family. Here’s the size-by-size breakdown:
- Qwen 2.5 0.5B / 1.5B / 3B — trivial. Loads in seconds, leaves >25 GB free for a long context window or a second model.
- Qwen 2.5 7B — comfortable at FP16 (14 GB) with a 32K context. Plenty of headroom.
- Qwen 2.5 14B — the production sweet spot. FP16 weights are 28 GB; the 5090’s 32 GB just fits with a 4-8K context. For longer contexts or bigger batches, drop to FP8 (~14 GB) and you have plenty of room.
- Qwen 2.5 32B — too large for FP16 (64 GB needed). Runs comfortably at INT4 (~24 GB). For FP16 you need a 6000 Pro or multi-GPU cluster.
- Qwen 2.5 72B — does not fit even at INT4 with reasonable KV cache (~38 GB needed). Skip the single 5090 for this size.
Throughput on a 5090
Indicative numbers from our vLLM benchmark suite:
- Qwen 2.5 7B FP16 — ~1,100 tok/s aggregate, ~85 tok/s single-stream
- Qwen 2.5 14B FP16 — ~720 tok/s aggregate, ~58 tok/s single-stream
- Qwen 2.5 14B FP8 — ~1,150 tok/s aggregate, ~80 tok/s single-stream
For more details see our best GPU for Qwen guide.
Frequently Asked Questions
The questions buyers actually ask before committing to a GPU server.
What’s the longest Qwen context I can run on a 5090?
Qwen 2.5 7B FP16 hits 64K context comfortably. Qwen 2.5 14B FP16 caps at ~16K with the default KV cache.
Should I quantise Qwen on a 5090?
For 14B+, FP8 is essentially free quality-wise and doubles your context budget. AWQ-INT4 is also fine for 7B/14B.
Can I run Qwen 2.5 32B at INT4 on a 5090?
Yes — fits with ~24 GB. Tight on context. The 6000 Pro is a better home for 32B.
Multi-GPU 5090 for Qwen 72B?
2× RTX 5090 (64 GB combined) handles Qwen 2.5 72B at INT4. See multi-GPU clusters.
Related Pages
Pages our visitors typically read next.
Ready to deploy?
Same-day deployment on in-stock GPUs. Talk to a specialist who actually understands your workload.
