GeForce RTX 5080 · Qwen3 30B-A3B · Q4_K_M
Only with CPU offload. Qwen3 30B-A3B at Q4_K_M does not fit in the 16 GB on GeForce RTX 5080, and a 32 GB system RAM pool runs it slowly, roughly 61–89 tokens per second.
| Piece | Value | Note |
|---|---|---|
| Weights | 17.3 GiB | Qwen3-30B-A3B-Q4_K_M.gguf |
| KV cache | 0.75 GiB | 8,192 tokens |
| Engine overhead | 0.50 GiB | Flat 0.5 GiB allowance |
| 5% headroom | 0.93 GiB | On the sum above |
| Total | 19.5 GiB | Offload |
| Fast memory | 16.0 GiB | 16 GB VRAM |
| Speed | 61–89 tok/s | Active-expert bandwidth range |
ollama run qwen3:30b-a3b
GeForce RTX 5080 · Qwen3 30B-A3B ·GGUF file · Methodology
FAQ
How was this fit calculated?
Weights 17.3 GiB, KV cache 0.75 GiB at 8,192 tokens, engine overhead 0.50 GiB, then 5% headroom, for 19.5 GiB. Fast memory available is 16.0 GiB.
What if I change the quant or the context?
This page is Q4_K_M at 8,192 tokens. A smaller quant or a shorter context lowers the total. The homepage calculator changes both without a new page.
Can I trust the tokens per second?
The range 61–89 tok/s is a bandwidth roofline, 55% to 80% efficient. This MoE model only reads the active experts, using an architecture-derived fraction of the file. Measure your own box before you buy around it.
What do I type?
ollama run qwen3:30b-a3b
Is GeForce RTX 5080 assumed to have extra RAM?
Yes. Discrete GPU pages assume 32 GB of system RAM. Offload uses whatever remains after 8 GB for Windows, and it is limited by the slower of DDR5-5600 and the PCIe link.