Radeon RX 9070 XT
Radeon RX 9070 XT has 16 GB of VRAM at 640 GB/s — about 14.88 GiB usable after driver and compositor overhead. 1811 of 2118 indexed models fit at 64K context with q4_0 KV.
What fits at 64K context
| Model | Best quant | Params | Weights● | KV● | Total in memory◐ | Headroom◐ | tok/s◐ |
|---|---|---|---|---|---|---|---|
| Llama3.2-24B-A3B-II-Dark-Champion-INSTRUCT-Heretic-Abliterated-UncensoredMoE | I1-Q5_K_M | 18.0B | 12.00 GiB | 1.97 GiB | 14.88 GiB | 0.00 GiB | 53±37% |
| GRM-2.6-Plus-0628 | Q3_K_S | 27.8B | 12.78 GiB | 1.13 GiB | 14.87 GiB | 0.01 GiB | 29±26.5% |
| ThinkingCap-Qwen3.6-27B | Q3_K_S | 27.4B | 12.78 GiB | 1.13 GiB | 14.87 GiB | 0.01 GiB | 29±26.5% |
| Tess-4-27B | Q3_K_S | 27.8B | 12.78 GiB | 1.13 GiB | 14.87 GiB | 0.01 GiB | 29±26.5% |
| Qwen3-16B-A3BMoE | Q6_K | 16.0B | 12.28 GiB | 1.69 GiB | 14.86 GiB | 0.02 GiB | 56±37% |
| Llama-3.2-3B | F16 | 3.2B | 11.98 GiB | 1.97 GiB | 14.86 GiB | 0.02 GiB | 29±26.5% |
| Gemma-4-31B-Isometry-RP | I1-IQ2_M | 32.7B | 10.73 GiB | 3.14 GiB | 14.85 GiB | 0.03 GiB | 29±26.5% |
| Gemma-4-Dark-Gemistry-31B | I1-IQ2_M | 32.7B | 10.73 GiB | 3.14 GiB | 14.85 GiB | 0.03 GiB | 29±26.5% |
| Prosopon-31B | I1-IQ2_M | 32.7B | 10.73 GiB | 3.14 GiB | 14.85 GiB | 0.03 GiB | 29±26.5% |
| Gemma-4-Novelist-Eclipse-31B | I1-IQ2_M | 32.7B | 10.73 GiB | 3.14 GiB | 14.85 GiB | 0.03 GiB | 29±26.5% |
| Giftige-Blume-31B-v1-StyleSwap | I1-IQ2_M | 32.7B | 10.73 GiB | 3.14 GiB | 14.85 GiB | 0.03 GiB | 29±26.5% |
| G4-MeroMero-31B-StyleSwap | I1-IQ2_M | 32.7B | 10.73 GiB | 3.14 GiB | 14.85 GiB | 0.03 GiB | 29±26.5% |
| Gemma-4-31B-StyleTune-heretic-ara | I1-IQ2_M | 32.7B | 10.73 GiB | 3.14 GiB | 14.85 GiB | 0.03 GiB | 29±26.5% |
| Pantheon-Reasoning-31B-1.1 | I1-IQ2_M | 32.7B | 10.73 GiB | 3.14 GiB | 14.85 GiB | 0.03 GiB | 29±26.5% |
| Gemma-4-31B-StyleTune | I1-IQ2_M | 32.7B | 10.73 GiB | 3.14 GiB | 14.85 GiB | 0.03 GiB | 29±26.5% |
| Barcenas-StyleTune-31B-Fable | I1-IQ2_M | 32.1B | 10.73 GiB | 3.14 GiB | 14.85 GiB | 0.03 GiB | 29±26.5% |
| Dolphin3.0-Llama3.2-3B | F32 | 3.2B | 11.98 GiB | 1.97 GiB | 14.85 GiB | 0.03 GiB | 29±26.5% |
| Hermes-3-Llama-3.2-3B | F32 | 3.2B | 11.98 GiB | 1.97 GiB | 14.85 GiB | 0.03 GiB | 29±26.5% |
| Tiger-Gemma-12B-v3 | Q8_0 | 12.8B | 12.65 GiB | 1.26 GiB | 14.85 GiB | 0.03 GiB | 29±26.5% |
| Goetia-26B-A4B-v1.4MoE | I1-Q3_K_L | 26.0B | 13.17 GiB | 0.79 GiB | 14.84 GiB | 0.04 GiB | 29±26.5% |
| G4-Moonlight-Dusk-26B-A4B-hereticMoE | I1-Q3_K_L | 26.5B | 13.17 GiB | 0.79 GiB | 14.84 GiB | 0.04 GiB | 29±26.5% |
| Pantheon-Reasoning-26B-A4B-1.1-hereticMoE | I1-Q3_K_L | 26.5B | 13.17 GiB | 0.79 GiB | 14.84 GiB | 0.04 GiB | 29±26.5% |
| G4-Moonlight-Dusk-26B-A4BMoE | I1-Q3_K_L | 26.5B | 13.17 GiB | 0.79 GiB | 14.84 GiB | 0.04 GiB | 29±26.5% |
| Chimera-X-26B-A4BMoE | I1-Q3_K_L | 26.5B | 13.17 GiB | 0.79 GiB | 14.84 GiB | 0.04 GiB | 29±26.5% |
| Pantheon-Reasoning-26B-A4B-1.1MoE | I1-Q3_K_L | 26.5B | 13.17 GiB | 0.79 GiB | 14.84 GiB | 0.04 GiB | 29±26.5% |
| Gemma-4-26B-A4B-StyleTune-V2MoE | I1-Q3_K_L | 26.5B | 13.17 GiB | 0.79 GiB | 14.84 GiB | 0.04 GiB | 29±26.5% |
| Gemma-4-26B-A4B-StyleTuneMoE | I1-Q3_K_L | 26.5B | 13.17 GiB | 0.79 GiB | 14.84 GiB | 0.04 GiB | 29±26.5% |
| gemma-4-26b-a4b-heretic-styletune-v2-headMoE | I1-Q3_K_L | 25.8B | 13.17 GiB | 0.79 GiB | 14.84 GiB | 0.04 GiB | 29±26.5% |
| Qwen3.6-VL-REAP-26B-A3BMoE | IQ4_XS | 26.6B | 13.59 GiB | 0.35 GiB | 14.84 GiB | 0.04 GiB | 116±37% |
| granite-20b-code-instruct-8k | Q5_K_L | 20.1B | 13.86 GiB | 0.00 GiB | 14.84 GiB | 0.04 GiB | 29±26.5% |
| Gemma4-Gutenberg-31B | IQ2_XS | 31.3B | 10.71 GiB | 3.14 GiB | 14.83 GiB | 0.05 GiB | 29±26.5% |
| gemma-4-31B-it | IQ2_XS | 31.3B | 10.71 GiB | 3.14 GiB | 14.83 GiB | 0.05 GiB | 29±26.5% |
| Gemma4-Gutenberg-31B-Heretic | IQ2_XS | 31.3B | 10.71 GiB | 3.14 GiB | 14.83 GiB | 0.05 GiB | 29±26.5% |
| Equinox-31B | IQ2_XS | 31.3B | 10.71 GiB | 3.14 GiB | 14.83 GiB | 0.05 GiB | 29±26.5% |
| gemma-4-31B-it-SDFT-Heretic-RP | IQ2_XS | 30.7B | 10.71 GiB | 3.14 GiB | 14.83 GiB | 0.05 GiB | 29±26.5% |
| ERNIE-4.5-21B-A3B-Thinking | Q4_1 | 21.8B | 12.93 GiB | 0.98 GiB | 14.83 GiB | 0.05 GiB | 29±26.5% |
| ERNIE-4.5-21B-A3B-PT | Q4_1 | 21.9B | 12.93 GiB | 0.98 GiB | 14.83 GiB | 0.05 GiB | 29±26.5% |
| Phi-3.5-MoE-instructMoEKV unresolved | IQ2_S | 41.9B | 11.67 GiB | 2.25 GiB | 14.83 GiB | 0.05 GiB | 51±37% |
| spoomplesmaxx-v2.1-30B | I1-Q2_K_S | 28.9B | 9.31 GiB | 4.50 GiB | 14.82 GiB | 0.06 GiB | 29±26.5% |
| Huihui-granite-4.1-30b-abliterated | I1-Q2_K_S | 28.9B | 9.31 GiB | 4.50 GiB | 14.82 GiB | 0.06 GiB | 29±26.5% |
| granite-4.1-30b-heretic | I1-Q2_K_S | 28.9B | 9.31 GiB | 4.50 GiB | 14.82 GiB | 0.06 GiB | 29±26.5% |
| GLM-4.7-Flash-REAP-23B-A3B-absolute-heresyMoE | I1-Q4_K_M | 23.0B | 12.98 GiB | 0.93 GiB | 14.82 GiB | 0.06 GiB | 77±37% |
| grug-27b | IQ3_M | 27.4B | 12.73 GiB | 1.13 GiB | 14.81 GiB | 0.07 GiB | 29±26.5% |
| Carnice-V2-27b | IQ3_M | 27.4B | 12.73 GiB | 1.13 GiB | 14.81 GiB | 0.07 GiB | 29±26.5% |
| Fara1.5-27B | IQ3_M | 27.4B | 12.73 GiB | 1.13 GiB | 14.81 GiB | 0.07 GiB | 29±26.5% |
| Olmo-3.1-32B-Instruct | IQ3_XS | 32.2B | 12.45 GiB | 1.36 GiB | 14.81 GiB | 0.07 GiB | 29±26.5% |
| Olmo-3.1-32B-Think | IQ3_XS | 32.2B | 12.45 GiB | 1.36 GiB | 14.81 GiB | 0.07 GiB | 29±26.5% |
| Olmo-3-32B-Think | IQ3_XS | 32.2B | 12.45 GiB | 1.36 GiB | 14.81 GiB | 0.07 GiB | 29±26.5% |
| Llama3.2-30B-A3B-II-Dark-Champion-INSTRUCT-Heretic-Abliterated-UncensoredMoE | I1-Q2_K | 30.0B | 10.60 GiB | 3.30 GiB | 14.81 GiB | 0.07 GiB | 43±37% |
| Yi-34B-200K-DARE-megamerge-v8 | I1-IQ2_XS | 34.4B | 9.60 GiB | 4.22 GiB | 14.80 GiB | 0.08 GiB | 29±26.5% |
| dolphin-2.9.1-yi-1.5-34b | I1-IQ2_XS | 34.4B | 9.60 GiB | 4.22 GiB | 14.80 GiB | 0.08 GiB | 29±26.5% |
| OrionStar-Yi-34B-Chat-Llama | I1-IQ2_XS | 34.4B | 9.60 GiB | 4.22 GiB | 14.80 GiB | 0.08 GiB | 29±26.5% |
| Yi-34B-200K-Llamafied | I1-IQ2_XS | 34.4B | 9.60 GiB | 4.22 GiB | 14.80 GiB | 0.08 GiB | 29±26.5% |
| Yi-1.5-34B | IQ2_XS | 34.4B | 9.60 GiB | 4.22 GiB | 14.80 GiB | 0.08 GiB | 29±26.5% |
| Merged-RP-Stew-V2-34B | I1-IQ2_XS | 34.4B | 9.60 GiB | 4.22 GiB | 14.80 GiB | 0.08 GiB | 29±26.5% |
| Capybara-Tess-Yi-34B-200K | I1-IQ2_XS | 34.4B | 9.60 GiB | 4.22 GiB | 14.80 GiB | 0.08 GiB | 29±26.5% |
| Gemma-3-27B-MeditronFO | I1-IQ3_M | 28.8B | 12.25 GiB | 1.58 GiB | 14.80 GiB | 0.08 GiB | 29±26.5% |
| GLM-4-32B-0414-Korean-Culture | I1-IQ3_XS | 32.6B | 12.72 GiB | 1.07 GiB | 14.78 GiB | 0.10 GiB | 29±26.5% |
| GLM-Z1-32B-0414 | IQ3_XS | 32.6B | 12.72 GiB | 1.07 GiB | 14.78 GiB | 0.10 GiB | 29±26.5% |
| GLM-4-32B-0414 | IQ3_XS | 32.6B | 12.72 GiB | 1.07 GiB | 14.78 GiB | 0.10 GiB | 29±26.5% |
Speed is modeled, not measured: decode is memory-bandwidth bound, so tokens per second is bytes read per token against achievable bandwidth. Mixture-of-experts models carry a wider band because only the routed experts are read each step, and few have been measured publicly.
Measured on this card
| Workload◍ | Median | Middle 50% | Runs |
|---|---|---|---|
| Image generation | 5.76 it/s | 2.54–11.88 | 118 |
| Prompt processing | 4275.83 tok/s | 3591.41–4809.87 | 30 |
| Text generation | 95.24 tok/s | 86.48–108.01 | 30 |
Aggregated from community-submitted runs, so the spread is wide by nature — it covers different models, resolutions, step counts and settings, not one controlled configuration. Read the middle 50% rather than the median alone. These figures are reproduced with attribution from vladmandic-sd-data-benchmark, which publishes no licence — so we display and link rather than redistribute them.
Questions people ask
- What AI models can a Radeon RX 9070 XT run?
- 1811 of 2118 indexed open-weight models fit a Radeon RX 9070 XT at 65,536 context with q4_0 KV cache, the largest being Llama3.2-24B-A3B-II-Dark-Champion-INSTRUCT-Heretic-Abliterated-Uncensored at I1-Q5_K_M. That covers text, vision-language, image, video and speech models.
- How much usable memory does a Radeon RX 9070 XT actually have?
- Its nameplate is 16 GB, but about 14.88 GiB is available to a model once driver and compositor overhead is accounted for.
- Is a Radeon RX 9070 XT fast for local AI?
- Its memory bandwidth is 640 GB/s, and that figure — not teraflops — is what governs token generation speed. Capacity decides what you can run; bandwidth decides how fast it runs.