AMD · consumer

Radeon RX 7650 GRE

Radeon RX 7650 GRE has 8 GB of VRAM at 288 GB/s — about 7.44 GiB usable after driver and compositor overhead. 1207 of 2118 indexed models fit at 16K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
8 GB
GDDR6
Bandwidth
288 GB/s
128-bit bus
Tensor FP16
—
dense
TDP
170 W
$249 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1019vision language 96embedding 26audio asr 38video 7image 1audio tts 20

What fits at 16K context

largest quantization that fits, per model · 1207 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Ministral-3-8B-Instruct-2512-BF16-abliteratedI1-IQ4_XS8.9B4.37 GiB2.13 GiB7.44 GiB0.00 GiB28±26.5%
Ministral-3-8B-Instruct-2512-BF16IQ4_XS8.9B4.37 GiB2.13 GiB7.44 GiB0.00 GiB28±26.5%
Amaretto-8BI1-IQ4_XS8.9B4.37 GiB2.13 GiB7.44 GiB0.00 GiB28±26.5%
zeta-2.1I1-IQ4_NL8.3B4.50 GiB2.00 GiB7.44 GiB0.00 GiB28±26.5%
Nexa-AI-4x4B-InstructMoEI1-Q2_K12.1B4.28 GiB2.25 GiB7.44 GiB0.00 GiB23±37%
llm-jp-4-8b-instructIQ4_XS8.6B4.49 GiB2.00 GiB7.43 GiB0.01 GiB28±26.5%
next-8bI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Supertron2-Reranker-8BI1-IQ4_XS8.8B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
next-ocrI1-IQ4_XS8.8B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Qwen3-VL-8B-GLM-4.7-Flash-Heretic-Uncensored-ThinkingI1-IQ4_XS8.8B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Midas-FableAgent-8BI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Qwen3-VL-8B-Heretic-1.3.0I1-IQ4_XS8.8B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Qwen3-VL-8B-Thinking-Unredacted-MAXI1-IQ4_XS8.8B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Qwen3-VL-8B-Instruct-Minecraft-MT-en-zhI1-IQ4_XS8.8B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Qwen-3-VL-8B-Instruct-hereticI1-IQ4_XS8.8B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Poe-8B-GLM5-Opus4.6-Sonnet4.5-Kimi-Grok-Gemini-3-pro-preview-HERETICI1-IQ4_XS8.8B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
ToolCUA-8BI1-IQ4_XS8.8B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Huihui-Qwen3-VL-8B-Instruct-abliteratedI1-IQ4_XS8.8B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Qwen3-VL-Reranker-8BI1-IQ4_XS8.8B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Salience-1-9BI1-IQ4_XS8.8B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Qwen3-VL-8B-Instruct-Uncensored-V2I1-IQ4_XS8.8B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Maestro1-9BI1-IQ4_XS8.8B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
GRaPE-2-FlashI1-IQ4_XS8.8B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Jan-v2-VL-medI1-IQ4_XS8.8B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Parable-Qwen3-8B-Claude-Fable-5I1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
ReasonCritic-7BI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
mythos-9b-unhinged-hereticI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Finch-8B-KTOI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Finch-8BI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
MathSmith-hc-Qwen3-8BI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
MiroThinker-v1.0-8BI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
mythos-9b-unhingedI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Ektome-Qwen3-8B-PristinelyUncensoredI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Marco-DeepResearch-8BI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
mythos-9b-mergedI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
qwen3-8b-apostateI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Josiefied-Qwen3-8B-abliterated-v1I1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
tmax-8bI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Qwen3-8B-DeepSeek-v3.2-Speciale-DistillIQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Qwen3-8B-abliteratedI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
story_generation_Qwen3_8B_RLI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
AReaL-boba-2-8B-OpenI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
DS-R1-Qwen3-8B-ArliAI-RpR-v4-SmallI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Nemotron-Orchestrator-8BIQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
S1-Base-8BI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Huihui-Qwen3-8B-abliterated-v2I1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Step3-VL-10B-BaseI1-IQ4_XS10.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
phi-2Q4_K_S2.8B1.52 GiB5.00 GiB7.43 GiB0.01 GiB28±26.5%
Qwen3-Reranker-8BI1-IQ4_XS8.2B4.25 GiB2.25 GiB7.43 GiB0.01 GiB28±26.5%
Aya-Medikal-V2I1-IQ4_NL8.0B4.48 GiB2.00 GiB7.43 GiB0.01 GiB28±26.5%
Kimi-VL-A3B-InstructMoEI1-IQ2_M16.4B6.03 GiB0.47 GiB7.42 GiB0.02 GiB70±37%
Moonlight-16B-A3B-InstructMoEIQ2_M16.0B6.03 GiB0.47 GiB7.42 GiB0.02 GiB70±37%
Phi-4-reasoningUD-IQ1_S14.7B3.33 GiB3.13 GiB7.42 GiB0.02 GiB28±26.5%
Phi-4-reasoning-plusUD-IQ1_S14.7B3.33 GiB3.13 GiB7.42 GiB0.02 GiB28±26.5%
Apriel-1.6-15b-ThinkerI1-IQ1_M14.9B3.47 GiB3.00 GiB7.42 GiB0.02 GiB28±26.5%
Wan2.1-T2V-1.3BQ4_01.4B6.50 GiB0.00 GiB7.42 GiB0.02 GiB28±26.5%
AMALIA-9B-0626-DPOQ3_K_S9.2B3.86 GiB2.63 GiB7.41 GiB0.03 GiB28±26.5%
Gemma-4-E4B-LuchadorQ5_K_L8.0B6.21 GiB0.29 GiB7.41 GiB0.03 GiB28±26.5%
Hy-MT2-7BQ4_18.0B4.47 GiB2.00 GiB7.41 GiB0.03 GiB28±26.5%
HomunculusIQ2_S12.5B3.96 GiB2.50 GiB7.41 GiB0.03 GiB28±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 7650 GRE run?
1207 of 2118 indexed open-weight models fit a Radeon RX 7650 GRE at 16,384 context with f16 KV cache, the largest being Ministral-3-8B-Instruct-2512-BF16-abliterated at I1-IQ4_XS. That covers text, vision-language, image, video and speech models.
How much usable memory does a Radeon RX 7650 GRE actually have?
Its nameplate is 8 GB, but about 7.44 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Radeon RX 7650 GRE fast for local AI?
Its memory bandwidth is 288 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.