Can the Radeon RX 9070 XT run Qwen2.5 Coder 32B Instruct?
Only with CPU offload at Q4_K_M.
Qwen2.5 Coder 32B Instruct is 32.5B parameters. At Q4_K_M the weights are about 18.2 GiB; with the 1.5 GiB runtime reserve that is 19.7 GiB against the card's 16GB, about 3.7 GiB more than the card has.
All three quants
16GB card| Quant | Weights | With reserve | Verdict | Headroom |
|---|---|---|---|---|
| Q4_K_Mthe everyday quant | 18.2 GiB | 19.7 GiB | Only with CPU offload | short 3.7 GiB |
| Q8_0near-lossless | 32.1 GiB | 33.6 GiB | Does not fit | short 17.6 GiB |
| FP16full weights | 60.5 GiB | 62.0 GiB | Does not fit | short 46.0 GiB |
Verdict rules: fits comfortably when weights plus reserve sit within 85% of VRAM; tight when they fit with little left for context; CPU offload when they exceed VRAM by up to 2x (runs, slowly, with layers in system RAM); no beyond that. The KV cache per token depends on the model's architecture and is not modelled here.
Not this card. These fit it comfortably at Q4.
- GeForce RTX 5090 32GB · headroom 12.3 GiB
- GeForce RTX 3090 (used) 24GB · headroom 4.3 GiB
- GeForce RTX 4090 (used) 24GB · headroom 4.3 GiB
- Radeon RX 7900 XTX 24GB · headroom 4.3 GiB
What the Radeon RX 9070 XT does run well → · Qwen2.5 Coder 32B Instruct on every card →
About the model
Qwen2.5 Coder 32B Instruct by Alibaba Qwen: 32,768-token native context, text, released 2024 under the Apache-2.0. Model card → · Hosted pricing on PicksByModel →