Radeon RX 9070 XT 16GB (desktop) for Local LLMs: VRAM and Model Fit
16 GB GDDR6 is discrete VRAM: 17,179,869,184 bytes (16.00 GiB). The default system reserve is 2,147,483,648 bytes (2.00 GiB), and the net is 15,032,385,536 bytes (14.00 GiB).
RX 9070 XT VRAM is 16 GB GDDR6. The model budget here is 14 GiB. The largest model in the table below that runs cleanly at 8,192 tokens is Qwen3-14B (Q5_K_M, 11.2 GiB–11.8 GiB).
| Model | Best file that runs | Total at 8,192 | Result | At 32,768 |
|---|---|---|---|---|
| Llama 3.1 8B Instruct | none | n/a | No | n/a |
| Qwen3-8B | Q8_0 | 9.4 GiB–10.0 GiB | Fits | Tight |
| Gemma 3 12B IT | Q4_0 | 8.5 GiB–9.0 GiB | Comfortable | Fits |
| Qwen3-14B | Q5_K_M | 11.2 GiB–11.8 GiB | Fits | No |
| gpt-oss-20b | MXFP4 | 13.6 GiB–15.5 GiB | Tight | No |
| Mistral-Small-24B-Instruct-2501 | none | n/a | No | n/a |
| Gemma 3 27B IT | none | n/a | No | n/a |
| Qwen3-30B-A3B | none | n/a | No | n/a |
| Qwen3-32B | none | n/a | No | n/a |
One sequence, fp16 KV cache, typical extras, the largest official file that fits. Formula estimates (see methodology), not runs on this card, and no speed figure is given. "n/a" means the model's native window is shorter than 32,768 tokens or no file runs. Cards with the same net budget are ranked on the best-GPU page.
Radeon RX 9070 XT 16GB (desktop): 16.00 GiB gross, default reserve 2.00 GiB, net calculator budget 14.00 GiB. Vendor-marketed GB for discrete GPU VRAM is treated as GiB (2^30 bytes), matching physical GDDR/HBM module organization. This is an explicit unit assumption.
This result is an estimate, not a measured trace.Discrete VRAM
Models for this device
Gross vs reserve vs net budget
- AMDAMD Radeon RX 9070 XT 16GB · desktop-gpu
- Memory kinddiscrete-vram
- Marketed16 GB GDDR6
- Gross17,179,869,184 bytes = 16.00 GiB
- Default reserve2,147,483,648 bytes = 2.00 GiB
- Net budget15,032,385,536 bytes = 14.00 GiB
The reserve is a configurable calculator assumption, not a measured desktop tax. Headless systems can lower it after measuring baseline board usage.
Catalog fit at 8,192 tokens
Each row is the same automatic pick as the calculator at 8,192 tokens: the highest precision at or above 3 bits per weight that fits. A file under that floor stays available for a manual choice and is not the automatic row. Concurrent sequences = 1, FP16 KV. Fit and software support are separate. Format compatibility is not a verified run.
Likely fits (43)
| Model | Quantization | Estimate | Fit | Software compatibility | Evidence |
|---|---|---|---|---|---|
| Qwen2.5-0.5B-Instruct | GGUF Q8_0 (Qwen) gguf-q8_0 · 8192 tokens | 0.85 GiB–1.24 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Gemma 3 1B IT | GGUF Q4_0 (Google QAT) gguf-q4_0 · 8192 tokens | 1.10 GiB–1.50 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen3-0.6B | GGUF Q8_0 (Qwen) gguf-q8_0 · 8192 tokens | 1.60 GiB–2.04 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen2.5-1.5B-Instruct | GGUF Q8_0 (Qwen) gguf-q8_0 · 8192 tokens | 2.11 GiB–2.53 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen2.5-Coder-1.5B-Instruct | GGUF Q8_0 (Qwen) gguf-q8_0 · 8192 tokens | 2.11 GiB–2.53 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Llama 3.2 1B Instruct | Official BF16 safetensors bf16 · 8192 tokens | 2.68 GiB–3.12 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen3-1.7B | GGUF Q8_0 gguf-q8_0 · 8192 tokens | 2.71 GiB–3.18 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen3-1.7B-Base | GGUF Q8_0 (mradermacher) gguf-q8_0 · 8192 tokens | 2.71 GiB–3.18 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Gemma 3 4B IT | GGUF Q4_0 (Google QAT) gguf-q4_0 · 8192 tokens | 3.33 GiB–3.79 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| SmolLM2-1.7B-Instruct | GGUF Q8_0 (Unsloth) gguf-q8_0 · 8192 tokens | 3.32 GiB–3.83 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen2.5-VL-3B-Instruct | GGUF Q8_0 (Unsloth) gguf-q8_0 · 8192 tokens | 3.47 GiB–3.92 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen2.5-3B-Instruct | GGUF Q8_0 gguf-q8_0 · 8192 tokens | 3.77 GiB–4.24 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen2.5-Coder-3B-Instruct | GGUF Q8_0 (Qwen) gguf-q8_0 · 8192 tokens | 3.77 GiB–4.24 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Phi-4-mini-instruct | GGUF Q8_0 (Unsloth) gguf-q8_0 · 8192 tokens | 4.93 GiB–5.45 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen3-4B | GGUF Q8_0 gguf-q8_0 · 8192 tokens | 5.24 GiB–5.77 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen3-4B-Instruct-2507 | GGUF Q8_0 (Unsloth) gguf-q8_0 · 8192 tokens | 5.24 GiB–5.77 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen3-VL-4B-Instruct | GGUF Q8_0 (Qwen) gguf-q8_0 · 8192 tokens | 5.24 GiB–5.77 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Llama 3.2 3B Instruct | Official BF16 safetensors bf16 · 8192 tokens | 6.98 GiB–7.54 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Phi-3.5-mini-instruct | GGUF Q8_0 (bartowski) gguf-q8_0 · 8192 tokens | 6.91 GiB–7.57 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen2-7B-Instruct | GGUF Q8_0 (Qwen) gguf-q8_0 · 8192 tokens | 8.10 GiB–8.66 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen2.5-VL-7B-Instruct | GGUF Q8_0 (Unsloth) gguf-q8_0 · 8192 tokens | 8.10 GiB–8.66 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| OpenMath-Nemotron-7B | GGUF Q8_0 (mradermacher) gguf-q8_0 · 8192 tokens | 8.10 GiB–8.66 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| DeepSeek-R1-Distill-Qwen-7B | GGUF Q8_0 (Unsloth) gguf-q8_0 · 8192 tokens | 8.10 GiB–8.66 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen2.5-7B-Instruct | GGUF Q8_0 gguf-q8_0 · 8192 tokens | 8.10 GiB–8.66 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen2.5-Coder-7B-Instruct | GGUF Q8_0 (Qwen) gguf-q8_0 · 8192 tokens | 8.10 GiB–8.66 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen2.5-Coder-7B | GGUF Q8_0 (mradermacher) gguf-q8_0 · 8192 tokens | 8.10 GiB–8.66 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Mistral-7B-Instruct-v0.3 | GGUF Q8_0 (bartowski) gguf-q8_0 · 8192 tokens | 8.30 GiB–8.89 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Gemma 3 12B IT | GGUF Q4_0 (Google QAT) gguf-q4_0 · 8192 tokens | 8.46 GiB–9.04 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| DeepSeek-R1-Distill-Llama-8B | GGUF Q8_0 (Unsloth) gguf-q8_0 · 8192 tokens | 9.08 GiB–9.68 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen2.5-Math-7B-Instruct | GGUF Q8_0 (bartowski) gguf-q8_0 · 8192 tokens | 8.10 GiB–9.96 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen3-8B | GGUF Q8_0 gguf-q8_0 · 8192 tokens | 9.36 GiB–9.98 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen3-8B-Base | GGUF Q8_0 (mradermacher) gguf-q8_0 · 8192 tokens | 9.36 GiB–9.98 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen3-VL-8B-Instruct | GGUF Q8_0 (Qwen) gguf-q8_0 · 8192 tokens | 9.36 GiB–9.98 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| granite-3.3-8b-instruct | GGUF Q8_0 (IBM) gguf-q8_0 · 8192 tokens | 9.46 GiB–10.09 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| GLM-4-9B-0414 | GGUF Q8_0 (Unsloth) gguf-q8_0 · 8192 tokens | 9.75 GiB–10.33 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Mistral-Nemo-Instruct-2407 | GGUF Q6_K (Unsloth) gguf-q6_k · 8192 tokens | 10.74 GiB–11.39 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen3-14B | GGUF Q5_K_M (Qwen) gguf-q5_k_m · 8192 tokens | 11.17 GiB–11.83 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| phi-4 | GGUF Q5_K_M (Unsloth) gguf-q5_k_m · 8192 tokens | 11.39 GiB–12.06 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| OpenMath-Nemotron-14B | GGUF Q5_K_M (mradermacher) gguf-q5_k_m · 8192 tokens | 11.41 GiB–12.09 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen2.5-Coder-14B-Instruct | GGUF Q5_K_M (Qwen) gguf-q5_k_m · 8192 tokens | 11.41 GiB–12.09 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Qwen2.5-14B-Instruct | GGUF Q5_K_M (Qwen) gguf-q5_k_m · 8192 tokens | 11.41 GiB–12.09 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| DeepSeek-V2-Lite-Chat | GGUF Q4_K_M (mradermacher) gguf-q4_k_m · 8192 tokens | 10.01 GiB–12.23 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
| Phi-4-reasoning | GGUF Q5_K_M (Unsloth) gguf-q5_k_m · 8192 tokens | 11.56 GiB–12.25 GiB | Likely fits | Not verified | This result is an estimate, not a measured trace. |
Tight (2)
| Model | Quantization | Estimate | Fit | Software compatibility | Evidence |
|---|---|---|---|---|---|
| Qwen3.8-27B | GGUF UD-IQ3_XXS (Unsloth) gguf-ud-iq3_xxs · 8192 tokens | 10.96 GiB–13.32 GiB | Tight | Not verified | This result is an estimate, not a measured trace. |
| gpt-oss-20b | MXFP4 safetensors mxfp4 · 8192 tokens | 13.13 GiB–13.78 GiB | Tight | Not verified | This result is an estimate, not a measured trace. |
Insufficient memory (41)
| Model | Quantization | Estimate | Fit | Software compatibility | Evidence |
|---|---|---|---|---|---|
| Mistral-Small-24B-Instruct-2501 | GGUF Q4_K_M (Unsloth) gguf-q4_k_m · 8192 tokens | 14.72 GiB–15.45 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Llama 3.1 8B Instruct | Official BF16 safetensors bf16 · 8192 tokens | 16.08 GiB–16.83 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Gemma 3 27B IT | GGUF Q4_0 (Google QAT) gguf-q4_0 · 8192 tokens | 17.20 GiB–17.97 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Qwen3-30B-A3B | GGUF Q4_K_M gguf-q4_k_m · 8192 tokens | 18.16 GiB–18.93 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Qwen3-30B-A3B-Instruct-2507 | GGUF Q4_K_M (Unsloth) gguf-q4_k_m · 8192 tokens | 18.16 GiB–18.93 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Qwen3-VL-30B-A3B-Instruct | GGUF Q4_K_M (Qwen) gguf-q4_k_m · 8192 tokens | 18.16 GiB–18.93 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Qwen3-32B | GGUF Q4_K_M gguf-q4_k_m · 8192 tokens | 20.53 GiB–21.41 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| OpenMath-Nemotron-32B | GGUF Q4_K_M (mradermacher) gguf-q4_k_m · 8192 tokens | 20.61 GiB–21.50 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Qwen2.5-VL-32B-Instruct | GGUF Q4_K_M (Unsloth) gguf-q4_k_m · 8192 tokens | 20.61 GiB–21.50 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| DeepSeek-R1-Distill-Qwen-32B | GGUF Q4_K_M (Unsloth) gguf-q4_k_m · 8192 tokens | 20.61 GiB–21.50 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| QwQ-32B | GGUF Q4_K_M (Qwen) gguf-q4_k_m · 8192 tokens | 20.61 GiB–21.50 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Qwen2.5-Coder-32B-Instruct | GGUF Q4_K_M (Qwen) gguf-q4_k_m · 8192 tokens | 20.61 GiB–21.50 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Qwen2.5-32B-Instruct | GGUF Q4_K_M gguf-q4_k_m · 8192 tokens | 20.61 GiB–21.50 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Drama 3 | Unofficial int4 W4A16 (GPTQ-Marlin, with codec) int4-w4a16-community · 8192 tokens | 22.67 GiB–27.03 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Mixtral-8x7B-Instruct-v0.1 | GGUF Q4_K_M (mradermacher) gguf-q4_k_m · 8192 tokens | 27.62 GiB–28.59 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Llama-3.1-Nemotron-70B-Instruct-HF | GGUF Q4_K_M (bartowski) gguf-q4_k_m · 8192 tokens | 42.23 GiB–43.57 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Qwen2.5-72B-Instruct | GGUF Q4_K_M gguf-q4_k_m · 8192 tokens | 43.61 GiB–44.98 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Qwen3-Next-80B-A3B-Instruct | GGUF Q4_K_M (Qwen) gguf-q4_k_m · 8192 tokens | 45.96 GiB–47.29 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| gpt-oss-120b | MXFP4 safetensors mxfp4 · 8192 tokens | 61.18 GiB–62.79 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| GLM-4.5-Air | GGUF Q4_K_M (Unsloth) gguf-q4_k_m · 8192 tokens | 69.53 GiB–71.36 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Qwen3.8-Flash-Next | GGUF Q4_K_M (bartowski) gguf-q4_k_m · 8192 tokens | 111.75 GiB–131.53 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| MiniMax-M2 | GGUF Q4_K_M (Unsloth) gguf-q4_k_m · 8192 tokens | 130.90 GiB–133.99 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Qwen3-235B-A22B | GGUF Q4_K_M (Qwen) gguf-q4_k_m · 8192 tokens | 133.99 GiB–137.11 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Qwen3-235B-A22B-Instruct-2507 | GGUF Q4_K_M (Unsloth) gguf-q4_k_m · 8192 tokens | 133.99 GiB–137.11 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Llama 3.3 70B Instruct | Official BF16 safetensors bf16 · 8192 tokens | 134.04 GiB–137.22 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| DeepSeek-V4-Flash-0731 | FP8 safetensors fp8 · 8192 tokens | 156.21 GiB–183.71 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| GLM-4.6 | GGUF Q4_K_M (Unsloth) gguf-q4_k_m · 8192 tokens | 203.78 GiB–208.37 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| GLM-4.5 | GGUF Q4_K_M (Unsloth) gguf-q4_k_m · 8192 tokens | 204.59 GiB–209.20 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| MiniMax-M3 | GGUF Q4_K_M (bartowski) gguf-q4_k_m · 8192 tokens | 244.40 GiB–249.71 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| GLM-5.3-Flash | GGUF Q8_0 (Unsloth) gguf-q8_0 · 8192 tokens | 318.04 GiB–373.58 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| DeepSeek-V3 | GGUF Q4_K_M (Unsloth) gguf-q4_k_m · 8192 tokens | 377.32 GiB–443.15 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| DeepSeek-R1 | GGUF Q4_K_M (Unsloth) gguf-q4_k_m · 8192 tokens | 377.32 GiB–443.15 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| DeepSeek-V3-0324 | GGUF Q4_K_M (Unsloth) gguf-q4_k_m · 8192 tokens | 377.32 GiB–443.15 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| DeepSeek-R1-0528 | GGUF Q4_K_M (Unsloth) gguf-q4_k_m · 8192 tokens | 377.79 GiB–443.71 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| DeepSeek-V3.1 | GGUF Q4_K_M (Unsloth) gguf-q4_k_m · 8192 tokens | 378.22 GiB–444.21 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Muse Spark Open | Official NVFP4 safetensors nvfp4 · 8192 tokens | 498.47 GiB–508.91 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| DeepSeek-V4.1-Flash | FP8 safetensors fp8 · 8192 tokens | 475.99 GiB–558.80 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Ling-3.1-flash | Official FP8 safetensors (F8_E4M3 experts + BF16 remainder) fp8 · 8192 tokens | 535.11 GiB–628.26 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Kimi-K2-Instruct | GGUF Q4_K_M (Unsloth) gguf-q4_k_m · 8192 tokens | 578.81 GiB–679.50 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| Kimi-K2-Thinking | GGUF Q4_K_M (Unsloth) gguf-q4_k_m · 8192 tokens | 579.24 GiB–680.01 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
| GLM-5.3 | GGUF Q8_0 (Unsloth) gguf-q8_0 · 8192 tokens | 747.13 GiB–876.98 GiB | Insufficient memory | Not verified | This result is an estimate, not a measured trace. |
Measured vs planned evidence on this SKU
This page is an estimate unless a matching runtime-resident measurement is attached.
Related pages
Sources
- AMD Radeon RX 9070 XT — Max Memory Size 16 GB; Memory Type GDDR6; Memory Interface 256-bit; Memory Bandwidth up to 640 GB/s; Typical Board Power 304 W
Calculator
Loading catalog snapshot…
Device and workload
Pick a GPU or unified-memory machine. Candidate models appear immediately.
Full-resident weight estimates use total parameters, including inactive MoE experts.
Pick a device to list candidate models. You do not need to choose a model first.