What LLMs Can the Radeon RX 9070 XT 16GB Run?
The Radeon RX 9070 XT 16GB has 16 GB of VRAM and 640 GB/s of bandwidth. It runs 28 of 58 models fully at Q4_K_M with 8K context, up to Gemma 4 26B-A4B, and generates about 64 tokens/s on an 8B model.
28 of 58 models fit
Radeon RX 9070 XT 16GB (16.0 GB) at Q4_K_M with 8K context. Largest that fits fully: Gemma 4 26B-A4B.
| Model | Needs | Status | Speed | Best quant that fits |
|---|---|---|---|---|
| gpt-oss-20b | 12.9 GB | Fits | ~112 tok/s | MXFP4 (official) |
| Qwen 2.5 14B | 10.7 GB | Fits | ~36 tok/s | Q6_K |
| DeepSeek R1 Distill Qwen 14B | 10.7 GB | Fits | ~36 tok/s | Q6_K |
| Phi-4 14B | 10.6 GB | Fits | ~36 tok/s | Q6_K |
| Qwen 3 14B | 10.5 GB | Fits | ~36 tok/s | Q6_K |
| Ministral 3 14B | 9.7 GB | Fits | ~39 tok/s | Q6_K |
| Mistral Nemo 12B | 9.0 GB | Fits | ~43 tok/s | Q8_0 |
| Gemma 3 12B | 8.8 GB | Fits | ~43 tok/s | Q8_0 |
| Ministral 3 8B | 6.7 GB | Fits | ~60 tok/s | Q8_0 |
| Qwen 3 8B | 6.6 GB | Fits | ~62 tok/s | Q8_0 |
| Qwen 3.5 9B | 6.4 GB | Fits | ~61 tok/s | Q8_0 |
| Llama 3.1 8B | 6.3 GB | Fits | ~64 tok/s | Q8_0 |
| DeepSeek R1 Distill Llama 8B | 6.3 GB | Fits | ~64 tok/s | Q8_0 |
| Mistral 7B v0.3 | 5.7 GB | Fits | ~70 tok/s | Q8_0 |
| Gemma 4 E4B | 5.7 GB | Fits | ~69 tok/s | Q8_0 |
| Qwen 2.5 7B | 5.6 GB | Fits | ~71 tok/s | Q8_0 |
| DeepSeek R1 Distill Qwen 7B | 5.6 GB | Fits | ~71 tok/s | Q8_0 |
| Qwen 3 4B | 4.2 GB | Fits | ~111 tok/s | FP16 / BF16 |
| Phi-4 Mini | 4.1 GB | Fits | ~118 tok/s | FP16 / BF16 |
| Gemma 4 E2B | 4.0 GB | Fits | ~108 tok/s | FP16 / BF16 |
| Gemma 3 4B | 3.7 GB | Fits | ~120 tok/s | FP16 / BF16 |
| Qwen 3.5 4B | 3.7 GB | Fits | ~124 tok/s | FP16 / BF16 |
| Ministral 3 3B | 3.5 GB | Fits | ~135 tok/s | FP16 / BF16 |
| Llama 3.2 3B | 3.5 GB | Fits | ~140 tok/s | FP16 / BF16 |
| Llama 3.2 1B | 1.7 GB | Fits | ~374 tok/s | FP16 / BF16 |
| Gemma 3 1B | 1.6 GB | Fits | ~485 tok/s | FP16 / BF16 |
| Mistral Small 3.1 24B | 15.6 GB | Tight fit | ~23 tok/s | IQ4_XS |
| Gemma 4 26B-A4B | 15.6 GB | Tight fit | ~96 tok/s | IQ4_XS |
| Qwen3-Next 80B-A3B | 46.3 GB | 33% on GPU | ~26 tok/s | – |
| Qwen 2.5 72B | 44.5 GB | 34% on GPU | ~1.3 tok/s | – |
| Llama 3.3 70B | 43.2 GB | 35% on GPU | ~1.4 tok/s | – |
| DeepSeek R1 Distill Llama 70B | 43.2 GB | 35% on GPU | ~1.4 tok/s | – |
| Mixtral 8x7B | 28.0 GB | 54% on GPU | ~9.3 tok/s | – |
| Qwen 2.5 32B | 21.4 GB | 70% on GPU | ~5.2 tok/s | Q2_K |
| Qwen 2.5 Coder 32B | 21.4 GB | 70% on GPU | ~5.2 tok/s | Q2_K |
| DeepSeek R1 Distill Qwen 32B | 21.4 GB | 70% on GPU | ~5.2 tok/s | Q2_K |
| Qwen 3 32B | 21.4 GB | 70% on GPU | ~5.2 tok/s | Q2_K |
| Qwen 3.5 35B-A3B | 21.0 GB | 72% on GPU | ~47 tok/s | Q2_K |
| Qwen 3.6 35B-A3B | 21.0 GB | 72% on GPU | ~47 tok/s | Q2_K |
| Gemma 4 31B | 19.8 GB | 76% on GPU | ~6.4 tok/s | Q2_K |
| Qwen 3 30B-A3B | 18.8 GB | 80% on GPU | ~46 tok/s | Q2_K |
| Qwen3-Coder 30B-A3B | 18.8 GB | 80% on GPU | ~46 tok/s | Q2_K |
| Gemma 3 27B | 17.6 GB | 86% on GPU | ~9.7 tok/s | Q3_K_M |
| Qwen 3.5 27B | 16.9 GB | 89% on GPU | ~12 tok/s | IQ4_XS |
| Qwen 3.6 27B | 16.9 GB | 89% on GPU | ~12 tok/s | IQ4_XS |
| Qwen 3.8 27B | 16.9 GB | 89% on GPU | ~12 tok/s | IQ4_XS |
| Kimi K2.6 | 566 GB | Too small | – | – |
| DeepSeek R1 671B | 380 GB | Too small | – | – |
| Llama 3.1 405B | 234 GB | Too small | – | – |
| Qwen 3.5 397B-A17B | 226 GB | Too small | – | – |
| DeepSeek V4 Flash | 161 GB | Too small | – | – |
| Qwen 3 235B-A22B | 135 GB | Too small | – | – |
| MiniMax-M2 | 133 GB | Too small | – | – |
| Qwen 3.5 122B-A10B | 70.2 GB | Too small | – | – |
| Mistral Small 4 119B | 68.2 GB | Too small | – | – |
| gpt-oss-120b | 67.2 GB | Too small | – | – |
| Llama 4 Scout | 64.0 GB | Too small | – | – |
| GLM-4.5-Air | 62.1 GB | Too small | – | – |
Radeon RX 9070 XT 16GB for local AI
AMD's RDNA 4 card with much stronger AI compute than RDNA 3. For LLM inference it sits close to an RTX 5070 Ti on models that fit in 16 GB.
Setup tips: supported by recent ROCm releases on Linux and by LM Studio's Vulkan and ROCm runtimes on Windows. Its bandwidth makes it quick for 14B models and gpt-oss-20b.
Among the 14 devices we track with about 16 GB of usable memory, the Radeon RX 9070 XT 16GB ranks 6 of 14 for text-generation speed, because generation speed follows memory bandwidth. Others in the same class: RTX 5080 16GB, RTX 5070 Ti 16GB, RTX 4080 Super 16GB, RTX 4080 16GB. They all hold the same models; the faster ones simply answer sooner.
Our picks for the Radeon RX 9070 XT 16GB
The model we would run first for each job, chosen from our recommendations by what fits fully with an 8K context. "Best quant" is the highest quality that fits; "max context" is at Q4_K_M or the native format.
| Task | Model | Best quant | Max context | Speed |
|---|---|---|---|---|
| Best all-rounder | Qwen 3 14B | Q6_K | 37K | ~36 tok/s |
| Coding | gpt-oss-20b | MXFP4 (official) | 55K | ~106 tok/s |
| Fastest replies | gpt-oss-20b | MXFP4 (official) | 55K | ~106 tok/s |
| Long documents (64K+) | Qwen 3.5 9B | Q8_0 | 256K | ~61 tok/s |
| Images in, text out | Ministral 3 14B | Q6_K | 42K | ~39 tok/s |
| Image generation | Qwen-Image | GGUF Q5_K_S | – | – |
| Spec | Value |
|---|---|
| Memory | 16 GB |
| Memory bandwidth | 640 GB/s |
| Architecture | RDNA 4 |
| Released | 2025-03 |
| Llama 3.1 8B Q4_K_M speed | ~64 tok/s |
| Largest model that fits | Gemma 4 26B-A4B |
Image and video models on the Radeon RX 9070 XT 16GB
Peak VRAM with model swapping (ComfyUI default on smaller cards).
| Model | FP8 | FP8 fit | GGUF Q4 | Q4 fit |
|---|---|---|---|---|
| Stable Diffusion 1.5 | 1.8 GB | Fits | 1.7 GB | Fits |
| SDXL 1.0 | 4.1 GB | Fits | 4.1 GB | Fits |
| Stable Diffusion 3.5 Medium | 6.4 GB | Fits | 6.4 GB | Fits |
| Stable Diffusion 3.5 Large | 8.6 GB | Fits | 6.4 GB | Fits |
| FLUX.1 [dev] | 12.1 GB | Fits | 7.3 GB | Fits |
| FLUX.1 [schnell] | 12.1 GB | Fits | 7.3 GB | Fits |
| Qwen-Image | 20.4 GB | 74% on GPU | 12.3 GB | Fits |
| Wan 2.2 A14B (T2V / I2V) | 18.8 GB | 80% on GPU | 13.2 GB | Fits |
| Wan 2.2 TI2V 5B | 8.0 GB | Fits | 6.5 GB | Fits |
| HunyuanVideo | 18.3 GB | 82% on GPU | 13.1 GB | Fits |
| FLUX.1 Kontext [dev] | 12.3 GB | Fits | 7.5 GB | Fits |
| FLUX.2 [dev] | 31.0 GB | 49% on GPU | 23.6 GB | 64% on GPU |
| Z-Image Turbo | 6.8 GB | Fits | 4.9 GB | Fits |
| Qwen-Image-Edit 2509 | 20.5 GB | 73% on GPU | 12.4 GB | Fits |
| LTX-2 | 19.2 GB | 78% on GPU | 12.6 GB | Fits |
Upgrade path
The next step up is the Radeon RX 7900 XT 20GB with 20 GB and 25% more bandwidth.
Comparisons
Radeon RX 9070 XT 16GB for LLMs: common questions
Can the Radeon RX 9070 XT 16GB run Llama 3.3 70B?
Only partly. It needs about 43.2 GB at Q4_K_M, so about 35% fits in VRAM and the rest runs from system RAM, which is slow.
Can the Radeon RX 9070 XT 16GB run Qwen 3 32B?
Only partly. It needs about 21.4 GB at Q4_K_M, so about 70% fits in VRAM and the rest runs from system RAM, which is slow. Q2_K fits fully.
Can the Radeon RX 9070 XT 16GB run Qwen 3 14B?
Yes. Qwen 3 14B needs about 10.5 GB at Q4_K_M with 8K context, and the best quality that fits is Q6_K.
How fast is the Radeon RX 9070 XT 16GB for LLMs?
About 64 tokens per second for an 8B model at Q4_K_M, based on its 640 GB/s memory bandwidth. Larger models scale down roughly in proportion to their size.