AMD · consumer

Radeon RX 9070 GRE

Radeon RX 9070 GRE has 12 GB of VRAM at 432 GB/s — about 11.16 GiB usable after driver and compositor overhead. 1568 of 2118 indexed models fit at 16K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
12 GB
GDDR6
Bandwidth
432 GB/s
192-bit bus
Tensor FP16
—
dense
TDP
220 W
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
vision language 130text 1336audio asr 39audio tts 21image 2video 14embedding 26

What fits at 16K context

largest quantization that fits, per model · 1568 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3.5-35B-A3BMoEIQ2_XXS36.0B9.94 GiB0.31 GiB11.16 GiB0.00 GiB111±37%
Qwen3.6-35B-A3BMoEIQ2_XXS36.0B9.94 GiB0.31 GiB11.16 GiB0.00 GiB111±37%
Fimbulvetr-11B-v2I1-Q5_K_M10.7B7.22 GiB3.00 GiB11.16 GiB0.00 GiB26±26.5%
Fallen-Gemma3-27B-v1IQ2_M27.4B8.84 GiB1.38 GiB11.16 GiB0.00 GiB26±26.5%
Kimi-Linear-48B-A3B-InstructMoEIQ1_S49.1B9.77 GiB0.47 GiB11.15 GiB0.01 GiB26±26.5%
Olmo-3.1-32B-InstructUD-IQ2_XXS32.2B8.30 GiB1.84 GiB11.14 GiB0.02 GiB27±26.5%
Olmo-3.1-32B-ThinkUD-IQ2_XXS32.2B8.30 GiB1.84 GiB11.14 GiB0.02 GiB27±26.5%
Olmo-3-32B-ThinkUD-IQ2_XXS32.2B8.30 GiB1.84 GiB11.14 GiB0.02 GiB27±26.5%
Carnice-Qwen3.6-MoE-35B-A3BMoEI1-IQ2_S36.0B9.92 GiB0.31 GiB11.14 GiB0.02 GiB111±37%
Qwen35B-Agent-R2-AbliteratedMoEI1-IQ2_S34.7B9.92 GiB0.31 GiB11.14 GiB0.02 GiB111±37%
Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-DistilledMoEI1-IQ2_S36.0B9.92 GiB0.31 GiB11.14 GiB0.02 GiB111±37%
Darwin-35B-A3B-OpusMoEI1-IQ2_S36.0B9.92 GiB0.31 GiB11.14 GiB0.02 GiB111±37%
Qwen35B-Agent-R2MoEI1-IQ2_S34.7B9.92 GiB0.31 GiB11.14 GiB0.02 GiB111±37%
Carnice-MoE-35B-A3BMoEI1-IQ2_S36.0B9.92 GiB0.31 GiB11.14 GiB0.02 GiB111±37%
spoomplesmaxx-flash-35B-A3MoEI1-IQ2_S35.1B9.92 GiB0.31 GiB11.14 GiB0.02 GiB111±37%
Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliteratedMoEI1-IQ2_S36.0B9.92 GiB0.31 GiB11.14 GiB0.02 GiB111±37%
Qwen3.6-35B-A3B-Uncensored-AggressiveMoEI1-IQ2_S35.1B9.92 GiB0.31 GiB11.14 GiB0.02 GiB111±37%
WorldSim-Opus-3.6-35B-A3BMoEI1-IQ2_S35.1B9.92 GiB0.31 GiB11.14 GiB0.02 GiB111±37%
Qwen3.6-35B-A3B-abliterated-MAXMoEI1-IQ2_S35.1B9.92 GiB0.31 GiB11.14 GiB0.02 GiB111±37%
Huihui-Qwen3.6-35B-A3B-abliteratedMoEI1-IQ2_S36.0B9.92 GiB0.31 GiB11.14 GiB0.02 GiB111±37%
Qwopus3.6-35B-A3B-v1MoEI1-IQ2_S36.0B9.92 GiB0.31 GiB11.14 GiB0.02 GiB111±37%
Qwen3.6-35B-A3B-StyleTuneMoEI1-IQ2_S35.1B9.92 GiB0.31 GiB11.14 GiB0.02 GiB111±37%
Qwen3.6-35B-A3B-abliteratedMoEI1-IQ2_S35.1B9.92 GiB0.31 GiB11.14 GiB0.02 GiB111±37%
Holo-3.1-35B-A3BMoEIQ2_S35.1B9.92 GiB0.31 GiB11.14 GiB0.02 GiB111±37%
0GM-1.0-35B-A3B-0427MoEI1-IQ2_S36.0B9.92 GiB0.31 GiB11.14 GiB0.02 GiB111±37%
Mistral-7B-v0.1KV unresolvedQ6_K7.2B8.20 GiB2.00 GiB11.14 GiB0.02 GiB27±26.5%
gemma-4-26B-A4B-itMoEUD-IQ2_M26.5B9.33 GiB0.92 GiB11.13 GiB0.03 GiB26±26.5%
NVIDIA-Nemotron-Nano-12B-v2IQ4_XS12.3B6.29 GiB3.88 GiB11.13 GiB0.03 GiB27±26.5%
Ministral-3-14B-Instruct-2512Q4_K_M13.9B7.67 GiB2.50 GiB11.13 GiB0.03 GiB27±26.5%
Ministral-3-14B-Reasoning-2512Q4_K_M13.9B7.67 GiB2.50 GiB11.13 GiB0.03 GiB27±26.5%
Ministral-3-14B-Instruct-2512-BF16-abliteratedI1-Q4_K_M13.9B7.67 GiB2.50 GiB11.13 GiB0.03 GiB27±26.5%
Ministral-3-14B-abliteratedQ4_K_M13.9B7.67 GiB2.50 GiB11.13 GiB0.03 GiB27±26.5%
Ministral-3-14B-Instruct-2512-BF16Q4_K_M13.9B7.67 GiB2.50 GiB11.13 GiB0.03 GiB27±26.5%
Ministral-3-14B-Reasoning-2512-UncensoredI1-Q4_K_M13.9B7.67 GiB2.50 GiB11.13 GiB0.03 GiB27±26.5%
pagestorm-research-preview-14b-full-bookQ4_K_M13.5B7.67 GiB2.50 GiB11.13 GiB0.03 GiB27±26.5%
Ministral-8B-Instruct-2410Q8_08.0B7.94 GiB2.25 GiB11.13 GiB0.03 GiB27±26.5%
Gemma-3-27B-MeditronFOI1-IQ2_XS28.8B8.29 GiB1.86 GiB11.12 GiB0.04 GiB27±26.5%
deepseek-coder-6.7B-kexerI1-IQ2_M6.7B2.20 GiB8.00 GiB11.12 GiB0.04 GiB27±26.5%
Magicoder-S-DS-6.7BI1-IQ2_M6.7B2.20 GiB8.00 GiB11.12 GiB0.04 GiB27±26.5%
deepseek-coder-6.7b-baseI1-IQ2_M6.7B2.20 GiB8.00 GiB11.12 GiB0.04 GiB27±26.5%
Apriel-1.6-15b-ThinkerI1-Q3_K_L14.9B7.18 GiB3.00 GiB11.12 GiB0.04 GiB27±26.5%
MathCoder2-CodeLlama-7BIQ2_M6.7B2.20 GiB8.00 GiB11.12 GiB0.04 GiB27±26.5%
WizardLM-7B-UncensoredI1-IQ2_M6.7B2.20 GiB8.00 GiB11.12 GiB0.04 GiB27±26.5%
Llama-2-7B-32K-InstructI1-IQ2_M6.7B2.20 GiB8.00 GiB11.12 GiB0.04 GiB27±26.5%
Luna-AI-Llama2-UncensoredI1-IQ2_M6.7B2.20 GiB8.00 GiB11.12 GiB0.04 GiB27±26.5%
Swallow-7b-NVE-instruct-hfI1-IQ2_M6.7B2.20 GiB8.00 GiB11.12 GiB0.04 GiB27±26.5%
zeta-2.1Q8_08.3B8.17 GiB2.00 GiB11.11 GiB0.05 GiB27±26.5%
zeta-2Q8_08.3B8.17 GiB2.00 GiB11.11 GiB0.05 GiB27±26.5%
UncensoredLM-DeepSeek-R1-Distill-Qwen-14BIQ4_XS14.2B7.29 GiB2.88 GiB11.11 GiB0.05 GiB27±26.5%
Nemotron-Mini-4B-InstructQ4_K_S4.2B8.19 GiB2.00 GiB11.11 GiB0.05 GiB27±26.5%
stable-code-3bF162.8B5.21 GiB5.00 GiB11.11 GiB0.05 GiB26±26.5%
InternVL3_5-30B-A3BQ2_K30.8B10.16 GiB0.00 GiB11.10 GiB0.06 GiB27±26.5%
Phi-3-medium-4k-instructIQ4_XS14.0B7.02 GiB3.13 GiB11.10 GiB0.06 GiB27±26.5%
Huihui-gemma-4-26B-A4B-it-abliteratedMoEUD-IQ2_M26.5B9.29 GiB0.92 GiB11.10 GiB0.06 GiB26±26.5%
Qwen3.6-35B-A3B-REAM-160-ru-agentMoEIQ3_S23.6B9.88 GiB0.31 GiB11.09 GiB0.07 GiB100±37%
Salience-1.5-FlashMoEIQ2_XS31.1B8.70 GiB1.50 GiB11.09 GiB0.07 GiB58±37%
GLM-4.7-Flash-REAP-23B-A3B-absolute-heresyMoEI1-IQ3_S23.0B9.35 GiB0.83 GiB11.09 GiB0.07 GiB68±37%
Qwen3.6-28BMoEI1-Q2_K28.2B9.87 GiB0.31 GiB11.09 GiB0.07 GiB106±37%
Qwen3.5-28BMoEI1-Q2_K28.7B9.87 GiB0.31 GiB11.09 GiB0.07 GiB106±37%
Ling-mini-2.0MoEQ4_116.3B9.57 GiB0.63 GiB11.09 GiB0.07 GiB93±37%
●From the file◐Predictedwhat these mean

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.

Questions people ask

What AI models can a Radeon RX 9070 GRE run?
1568 of 2118 indexed open-weight models fit a Radeon RX 9070 GRE at 16,384 context with f16 KV cache, the largest being Qwen3.5-35B-A3B at IQ2_XXS. That covers text, vision-language, image, video and speech models.
How much usable memory does a Radeon RX 9070 GRE actually have?
Its nameplate is 12 GB, but about 11.16 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Radeon RX 9070 GRE fast for local AI?
Its memory bandwidth is 432 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.