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. 974 of 2118 indexed models fit at 64K 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
text 782vision language 98audio asr 38audio tts 20video 14image 1embedding 21

What fits at 64K context

largest quantization that fits, per model · 974 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
ERNIE-4.5-21B-A3B-PTUD-IQ1_M21.9B6.74 GiB3.50 GiB11.16 GiB0.00 GiB26±26.5%
Gemma-4-12B-StyleTuneI1-IQ3_M13.0B5.74 GiB4.47 GiB11.16 GiB0.00 GiB26±26.5%
gemma-4-12b-heretic-styletune-headI1-IQ3_M12.0B5.74 GiB4.47 GiB11.16 GiB0.00 GiB26±26.5%
syrian-gemma-12bI1-IQ3_M13.0B5.74 GiB4.47 GiB11.16 GiB0.00 GiB26±26.5%
dolphincoder-starcoder2-15bKV unresolvedI1-IQ2_M16.0B5.16 GiB5.00 GiB11.16 GiB0.00 GiB27±26.5%
starcoder2-15bKV unresolvedIQ2_M16.0B5.16 GiB5.00 GiB11.16 GiB0.00 GiB27±26.5%
Apertus-8B-Instruct-2509UD-IQ1_M8.1B2.19 GiB8.00 GiB11.15 GiB0.01 GiB27±26.5%
SmolLM3-3BBF163.1B5.74 GiB4.50 GiB11.15 GiB0.01 GiB26±26.5%
GLM-4.6V-FlashQ6_K10.3B7.70 GiB2.50 GiB11.14 GiB0.02 GiB27±26.5%
GLM-Z1-9B-0414Q6_K9.4B7.70 GiB2.50 GiB11.14 GiB0.02 GiB27±26.5%
glm4.1v-9b-base-sftI1-Q6_K10.3B7.70 GiB2.50 GiB11.14 GiB0.02 GiB27±26.5%
GLM-4-9B-0414Q6_K9.4B7.70 GiB2.50 GiB11.14 GiB0.02 GiB27±26.5%
GLM-4.1V-9B-ThinkingQ6_K10.3B7.70 GiB2.50 GiB11.14 GiB0.02 GiB27±26.5%
Aya-Medikal-V2I1-IQ1_M8.0B2.19 GiB8.00 GiB11.13 GiB0.03 GiB27±26.5%
DeepSeek-Coder-V2-Lite-BaseMoEI1-Q4_015.7B8.32 GiB1.90 GiB11.13 GiB0.03 GiB48±37%
Qwen3-VL-2B-ThinkingBF162.1B3.21 GiB7.00 GiB11.10 GiB0.06 GiB26±26.5%
Qwen3-VL-Reranker-2BF162.1B3.21 GiB7.00 GiB11.10 GiB0.06 GiB26±26.5%
Qwen3-VL-Embedding-2BF162.1B3.21 GiB7.00 GiB11.10 GiB0.06 GiB26±26.5%
Qwen3-VL-2B-InstructBF162.1B3.21 GiB7.00 GiB11.10 GiB0.06 GiB26±26.5%
Atomight-V2.5-1.7BF161.7B3.21 GiB7.00 GiB11.10 GiB0.06 GiB26±26.5%
OpenCaption-2B-VL-SFT-v1.0F162.1B3.21 GiB7.00 GiB11.10 GiB0.06 GiB26±26.5%
OpenClaude-1.7B-MergedBF161.7B3.21 GiB7.00 GiB11.10 GiB0.06 GiB26±26.5%
gaon-1.7b-v2-instructF161.7B3.21 GiB7.00 GiB11.10 GiB0.06 GiB26±26.5%
gaon-1.7b-v2-translateF161.7B3.21 GiB7.00 GiB11.10 GiB0.06 GiB26±26.5%
Lightning-1.7BBF161.7B3.21 GiB7.00 GiB11.10 GiB0.06 GiB26±26.5%
InternVL3_5-30B-A3BQ2_K30.8B10.16 GiB0.00 GiB11.10 GiB0.06 GiB27±26.5%
Ternary-Bonsai-1.7B-unpackedF161.7B3.21 GiB7.00 GiB11.10 GiB0.06 GiB26±26.5%
DeepSeek-Coder-V2-Lite-InstructMoEIQ4_NL15.7B8.29 GiB1.90 GiB11.10 GiB0.06 GiB48±37%
DeepSeek-V2-Lite-ChatMoEIQ4_NL15.7B8.29 GiB1.90 GiB11.10 GiB0.06 GiB48±37%
Falcon3-7B-InstructQ3_K_S7.5B3.14 GiB7.00 GiB11.10 GiB0.06 GiB27±26.5%
granite-speech-4.1-2b-plusF162.1B5.21 GiB5.00 GiB11.10 GiB0.06 GiB26±26.5%
Dolphin3.0-Llama3.2-3BQ8_03.2B3.19 GiB7.00 GiB11.10 GiB0.06 GiB27±26.5%
Llama-Song-Stream-3B-InstructQ8_03.2B3.19 GiB7.00 GiB11.10 GiB0.06 GiB27±26.5%
Llama-Doctor-3.2-3B-InstructQ8_03.2B3.19 GiB7.00 GiB11.10 GiB0.06 GiB27±26.5%
Llama-3.2-3B-Instruct-roleplay-tunedQ8_03.2B3.19 GiB7.00 GiB11.10 GiB0.06 GiB27±26.5%
llama-3.2-Korean-Bllossom-3BQ8_03.2B3.19 GiB7.00 GiB11.10 GiB0.06 GiB27±26.5%
Llama-3.2-3B-Instruct-uncensoredQ8_03.2B3.19 GiB7.00 GiB11.10 GiB0.06 GiB27±26.5%
Llama-3.2-3B-Instruct-heretic-ablitered-uncensoredQ8_03.2B3.19 GiB7.00 GiB11.10 GiB0.06 GiB27±26.5%
llama-3.2-3b-instruct-bnb-4bitQ8_03.3B3.19 GiB7.00 GiB11.10 GiB0.06 GiB27±26.5%
Llama-3.2-3B-InstructQ8_03.2B3.19 GiB7.00 GiB11.10 GiB0.06 GiB27±26.5%
Llama3.2-3B-creative-writer-v0.1Q8_03.2B3.19 GiB7.00 GiB11.10 GiB0.06 GiB27±26.5%
Firefly-V3.2Q8_03.2B3.19 GiB7.00 GiB11.10 GiB0.06 GiB27±26.5%
Firefly-V3Q8_03.2B3.19 GiB7.00 GiB11.10 GiB0.06 GiB27±26.5%
Hermes-3-Llama-3.2-3BQ8_03.2B3.19 GiB7.00 GiB11.10 GiB0.06 GiB27±26.5%
Llama-3.2-3BQ8_03.2B3.19 GiB7.00 GiB11.10 GiB0.06 GiB27±26.5%
ERNIE-21B-A3B-Thinking-Gemini-3-Pro-High-Reasoning-V2I1-IQ2_M21.8B6.68 GiB3.50 GiB11.10 GiB0.06 GiB27±26.5%
ERNIE-21B-A3B-Claude-4.5-High-OPUS-ThinkingI1-IQ2_M21.8B6.68 GiB3.50 GiB11.10 GiB0.06 GiB27±26.5%
ERNIE-4.5-21B-A3B-ThinkingI1-IQ2_M21.8B6.68 GiB3.50 GiB11.10 GiB0.06 GiB27±26.5%
dolphin-2.9.3-mistral-7B-32kI1-IQ2_S7.2B2.16 GiB8.00 GiB11.09 GiB0.07 GiB27±26.5%
Mistral-7B-v0.3IQ2_S7.2B2.16 GiB8.00 GiB11.09 GiB0.07 GiB27±26.5%
Mistral-7B-Instruct-v0.3-ParasiteI1-IQ2_S7.2B2.16 GiB8.00 GiB11.09 GiB0.07 GiB27±26.5%
Mistral-7B-Instruct-v0.3-JbliteratedI1-IQ2_S7.2B2.16 GiB8.00 GiB11.09 GiB0.07 GiB27±26.5%
Mathstral-7B-v0.1IQ2_S7.2B2.16 GiB8.00 GiB11.09 GiB0.07 GiB27±26.5%
AMD-OLMo-1B-SFT-DPOF161.2B2.19 GiB8.00 GiB11.09 GiB0.07 GiB27±26.5%
ZAYA1-8B-CoderMoEQ4_K_M8.8B5.19 GiB5.00 GiB11.09 GiB0.07 GiB27±26.5%
SciPhi-Self-RAG-Mistral-7B-32kKV unresolvedI1-IQ2_S7.2B2.15 GiB8.00 GiB11.09 GiB0.07 GiB27±26.5%
dolphin-2.2.1-mistral-7bKV unresolvedI1-IQ2_S7.2B2.15 GiB8.00 GiB11.09 GiB0.07 GiB27±26.5%
OpenChat-3.5-7B-Qwen-v2.0KV unresolvedI1-IQ2_S7.2B2.15 GiB8.00 GiB11.09 GiB0.07 GiB27±26.5%
openchat-3.5-0106KV unresolvedI1-IQ2_S7.2B2.15 GiB8.00 GiB11.09 GiB0.07 GiB27±26.5%
Mistral-7B-Instruct-v0.1KV unresolvedI1-IQ2_S7.2B2.15 GiB8.00 GiB11.09 GiB0.07 GiB27±26.5%
●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?
974 of 2118 indexed open-weight models fit a Radeon RX 9070 GRE at 65,536 context with f16 KV cache, the largest being ERNIE-4.5-21B-A3B-PT at UD-IQ1_M. 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.