Radeon RX 9070 XT
16 GB GDDR6 · 640 GB/s · 304 W
RX 9070 XT 16 GB · 64 GB RAM → gpt-oss-20b
GPU · AMD
RX 9070 XT for local LLMs: which models run?
With 16 GB of VRAM and 640 GB/s of bandwidth, the RX 9070 XT runs models up to about gpt-oss-20b entirely on the GPU.
- VRAM16 GB GDDR6
- Bandwidth640 GB/s
- Board power304 W
- ArchitectureRDNA 4
Specifications
- VRAM
- 16 GB
- Memory type
- GDDR6
- Memory bus
- 256-bit
- Bandwidth
- 640 GB/s
- Board power
- 304 W
- Backends
- ROCm, Vulkan
- Interface
- PCIe 5.0 x16
- Released
- March 2025
- Segment
- Consumer
- FP16 compute
- Not available
Official specData updated 24 days agoSource: AMD specifications
What the RX 9070 XT runs
| gpt-oss-20bMXFP4 · 8K~13 GB · 97–134 tok/s estimate | Excellent |
|---|---|
| Mistral Nemo 12B InstructQ6_K · 8K~12 GB · 36–52 tok/s estimate | Excellent |
| Gemma 2 9B ITQ8_0 · 8K~12.8 GB · 35–51 tok/s estimate | Excellent |
| Qwen3 8BQ8_0 · 8K~10.6 GB · 41–59 tok/s estimate | Excellent |
| Llama 3.1 8B InstructQ8_0 · 8K~10.2 GB · 43–61 tok/s estimate | Excellent |
| DeepSeek R1 Distill Llama 8BQ8_0 · 8K~10.2 GB · 43–61 tok/s estimate | Excellent |
| Qwen2.5 7B InstructQ8_0 · 8K~9.3 GB · 47–67 tok/s estimate | Excellent |
| Mistral 7B Instruct v0.3Q8_0 · 8K~9.2 GB · 47–67 tok/s estimate | Excellent |
| Gemma 3 4B ITQ8_0 · 8K~5.5 GB · 78–118 tok/s estimate | Excellent |
| Qwen3 4BQ8_0 · 8K~6.2 GB · 70–105 tok/s estimate | Excellent |
| Phi-4 Mini Instruct 3.8BQ8_0 · 8K~6 GB · 75–112 tok/s estimate | Excellent |
| Llama 3.2 3B InstructQ8_0 · 8K~5.2 GB · 88–132 tok/s estimate | Excellent |
| Qwen3 1.7BQ8_0 · 8K~3.7 GB · 127–203 tok/s estimate | Excellent |
| Llama 3.2 1B InstructQ8_0 · 8K~2.6 GB · 208–328 tok/s estimate | Excellent |
| Qwen3 0.6BQ8_0 · 8K~2.6 GB · 191–338 tok/s estimate | Excellent |
| Mistral Small 3.2 24B InstructIQ4_XS · 8K~14.7 GB · 30–42 tok/s estimate | Compatible |
| Qwen3 14BQ6_K · 8K~14.1 GB · 31–44 tok/s estimate | Compatible |
| DeepSeek R1 Distill Qwen 14BQ6_K · 8K~14.3 GB · 30–43 tok/s estimate | Compatible |
| Qwen2.5 14B InstructQ6_K · 8K~14.3 GB · 30–43 tok/s estimate | Compatible |
| Phi-4 14BQ6_K · 8K~14.2 GB · 31–44 tok/s estimate | Compatible |
| Gemma 3 12B ITQ8_0 · 8K~14.2 GB · 30–43 tok/s estimate | Compatible |
| Qwen3 30B-A3BQ3_K_M · 8K~15.4 GB · 75–136 tok/s estimate | Compatible with limits |
| Gemma 3 27B ITQ3_K_M · 8K~15.4 GB · 27–39 tok/s estimate | Compatible with limits |
| Mixtral 8x7B InstructQ3_K_M · 8K~21.6 GB · 9–18 tok/s estimate | CPU/GPU offload |
| Qwen3 32BQ3_K_M · 8K~18.6 GB · 5–10 tok/s estimate | CPU/GPU offload |
| DeepSeek R1 Distill Qwen 32BQ3_K_M · 8K~18.6 GB · 5.5–10 tok/s estimate | CPU/GPU offload |
| Qwen2.5 32B InstructQ3_K_M · 8K~18.6 GB · 5.5–10 tok/s estimate | CPU/GPU offload |
| Qwen2.5 Coder 32B InstructQ3_K_M · 8K~18.6 GB · 5.5–10 tok/s estimate | CPU/GPU offload |
| QwQ 32BQ3_K_M · 8K~18.6 GB · 5.5–10 tok/s estimate | CPU/GPU offload |
| Qwen3 Coder 30B-A3B InstructQ3_K_M · 8K~16 GB · 50–111 tok/s estimate | CPU/GPU offload |
| Gemma 2 27B ITQ3_K_M · 8K~16.5 GB · 12–21 tok/s estimate | CPU/GPU offload |
| gpt-oss-120bMXFP4 · 8K~63.3 GB · 8.5–17 tok/s estimate | Possible but slow |
| Llama 4 Scout 17B-16E InstructQ3_K_M · 8K~55.7 GB · 3–5.5 tok/s estimate | Possible but slow |
| Qwen2.5 72B InstructQ3_K_M · 8K~38.3 GB · 0.9–1.7 tok/s estimate | Not recommended |
| Llama 3.3 70B InstructQ3_K_M · 8K~37 GB · 0.9–1.8 tok/s estimate | Not recommended |
| DeepSeek R1 Distill Llama 70BQ3_K_M · 8K~37 GB · 0.9–1.8 tok/s estimate | Not recommended |
| DeepSeek R1 (671B)Q8_0 · 8K~699 GB · No data | Incompatible |
| Qwen3 235B-A22BQ8_0 · 8K~247 GB · No data | Incompatible |
For each model, the highest-precision quantization that runs entirely in fast memory; otherwise the one with the best result. Speeds are ranges; "estimate" unless a benchmark exists for this exact hardware.
EstimateData updated 24 days agoSource: Run LLM Locally engine
Limits in plain words
16 GB of VRAM holds the weights and KV cache of models up to about gpt-oss-20b at Q4_K_M. Bigger models spill into system RAM and slow down sharply: generation speed is then limited by RAM bandwidth, not by the GPU.
Where to buy a RX 9070 XT
Prices and links follow the market you choose.
No current offer imported for this market: the link opens a marketplace search.
Frequently asked questions
Which LLMs can the RX 9070 XT run?
Entirely on the GPU: Qwen3 30B-A3B (Q3_K_M), Gemma 3 27B IT (Q3_K_M), Mistral Small 3.2 24B Instruct (IQ4_XS), gpt-oss-20b (MXFP4), Qwen3 14B (Q6_K), DeepSeek R1 Distill Qwen 14B (Q6_K), Qwen2.5 14B Instruct (Q6_K), Phi-4 14B (Q6_K). Larger models need CPU/GPU offload.
Can the RX 9070 XT run a 70B model?
Not recommended: a 70B model at Q4_K_M needs about 45.1 GB; the RX 9070 XT has 16 GB.
How is the speed estimated?
Token generation reads every active weight once per token, so it is limited by memory bandwidth (640 GB/s here). We apply an efficiency range per backend and label the result as an estimate until real benchmarks exist.