AMD · unified x86

AMD Ryzen AI Max 390 (Radeon 8050S)

AMD Ryzen AI Max 390 (Radeon 8050S) has 32 GB of VRAM at 256 GB/s — about 22.32 GiB usable after driver and compositor overhead. 1998 of 2118 indexed models fit at 4K context with q4_0 KV. Note only 24 GB of its 32 GB is allocatable to the GPU.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
32 GB
LPDDR5X-8000
Bandwidth
256 GB/s
256-bit bus
Tensor FP16
—
dense
TDP
120 W
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1716vision language 178audio tts 21image 2video 16audio asr 39embedding 26

What fits at 4K context

largest quantization that fits, per model · 1998 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
NVIDIA-Nemotron-Nano-12B-v2BF1612.3B22.94 GiB0.27 GiB23.98 GiB0.02 GiB6±25%
Maenad-70BI1-Q2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-Q2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
Rombos-LLM-70b-Llama-3.3I1-Q2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
L3.3-Electra-R1-70bI1-Q2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
Latxa-Llama-3.1-70B-Instruct-v2I1-Q2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
Llama-3.3_70_b_uncensored_continuedI1-Q2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
Llama-3.3-70B-Instruct-abliteratedI1-Q2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
grok-oss-Revenant-70BI1-Q2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
Llama-3.1-Nemotron-70B-Instruct-HFI1-Q2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
L3.3-70B-Euryale-v2.3I1-Q2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
Hermes-4-70B-hereticI1-Q2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
Llama-3.1-70BQ2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
Golem-70B-v1bI1-Q2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-Q2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
DeepSeek-R1-Distill-Llama-70B-hereticI1-Q2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
Legion-V2.1-LLaMa-70BI1-Q2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
Assistant_Pepe_70BI1-Q2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
Athene-70BQ2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
Hermes-3-Llama-3.1-70BQ2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
L3.3-70B-Magnum-DiamondQ2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
Meta-Llama-3-70B-Instruct-abliterated-v3.5Q2_K_S70.6B22.79 GiB0.35 GiB23.97 GiB0.03 GiB6±25%
Laguna-S-2.1MoEIQ1_S118B23.15 GiB0.09 GiB23.96 GiB0.04 GiB33±37%
Qwen-AgentWorld-35B-A3BMoEUD-Q5_K_S34.7B23.23 GiB0.02 GiB23.96 GiB0.04 GiB36±37%
Ornith-1.0-35BMoEUD-Q5_K_S34.7B23.23 GiB0.02 GiB23.96 GiB0.04 GiB36±37%
WizardLM-Uncensored-SuperCOT-StoryTelling-30bQ5_K_M32.5B21.46 GiB1.71 GiB23.95 GiB0.05 GiB6±25%
Wizard-Vicuna-30B-UncensoredI1-Q5_K_M32.5B21.46 GiB1.71 GiB23.95 GiB0.05 GiB6±25%
archangel_sft-kto_llama30bI1-Q5_K_M32.5B21.46 GiB1.71 GiB23.95 GiB0.05 GiB6±25%
ALIA-40b-fc-2606I1-Q4_K_M40.4B22.90 GiB0.21 GiB23.93 GiB0.07 GiB6±25%
ALIA-40b-instruct-2606I1-Q4_K_M40.4B22.90 GiB0.21 GiB23.93 GiB0.07 GiB6±25%
Qwen3-Coder-Next-REAMMoEI1-IQ3_XS60.3B23.19 GiB0.03 GiB23.91 GiB0.09 GiB38±37%
Nemotron-Labs-Audex-30B-A3BQ4_K_M32.0B23.16 GiB0.00 GiB23.90 GiB0.10 GiB6±25%
Le-Chaton-Slim-23BMoEQ8_023.3B23.07 GiB0.11 GiB23.90 GiB0.10 GiB14±37%
Nemotron-Cascade-2-30B-A3BMoEQ4_K_L31.6B23.15 GiB0.06 GiB23.89 GiB0.11 GiB31±37%
Llama3.2-30B-A3B-II-Dark-Champion-INSTRUCT-Heretic-Abliterated-UncensoredMoEI1-Q6_K30.0B22.97 GiB0.21 GiB23.89 GiB0.11 GiB18±37%
Qwen3.5-122B-A10B-hereticMoEI1-IQ1_S123B23.13 GiB0.03 GiB23.89 GiB0.11 GiB36±37%
DeepSeek-R1-Distill-Llama-70BUD-IQ2_M70.6B22.70 GiB0.35 GiB23.87 GiB0.13 GiB6±25%
Qwen3.5-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-ThinkingI1-Q4_139.5B23.00 GiB0.11 GiB23.87 GiB0.13 GiB6±25%
Qwen3.5-40B-RoughHouse-Claude-4.6-Opus-Polar-Deckard-Uncensored-Heretic-ThinkingI1-Q4_139.5B23.00 GiB0.11 GiB23.87 GiB0.13 GiB6±25%
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16UD-Q5_K_S33.0B23.10 GiB0.00 GiB23.85 GiB0.15 GiB6±25%
Apertus-70B-Instruct-2509IQ2_M70.6B22.61 GiB0.35 GiB23.84 GiB0.16 GiB6±25%
CalmeRys-78B-Orpo-v0.1I1-IQ1_S78.0B22.62 GiB0.38 GiB23.83 GiB0.17 GiB6±25%
granite-34b-code-base-8kI1-Q5_K_M33.7B23.05 GiB0.00 GiB23.83 GiB0.17 GiB6±25%
Llama-4-Scout-17B-16E-Instruct-abliterated-v2MoEKV unresolvedI1-IQ1_M109B22.88 GiB0.21 GiB23.82 GiB0.18 GiB26±37%
Hy-MT2-30B-A3BMoEQ6_K30.1B23.01 GiB0.11 GiB23.81 GiB0.19 GiB27±37%
Hermes-4-70BUD-IQ2_M70.6B22.63 GiB0.35 GiB23.80 GiB0.20 GiB6±25%
Llama-3.3-70B-InstructUD-IQ2_M70.6B22.63 GiB0.35 GiB23.80 GiB0.20 GiB6±25%
Huihui-Qwen3.5-35B-A3B-abliteratedMoEI1-Q5_K_M36.0B23.06 GiB0.02 GiB23.79 GiB0.21 GiB36±37%
Qwen3.5-35B-A3B-BaseMoEI1-Q5_K_M36.0B23.06 GiB0.02 GiB23.79 GiB0.21 GiB36±37%
Qwen3.5-35B-A3B-ultra-uncensored-hereticMoEQ5_K_M35.1B23.06 GiB0.02 GiB23.79 GiB0.21 GiB36±37%
Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-DistilledMoEI1-Q5_K_M36.0B23.06 GiB0.02 GiB23.79 GiB0.21 GiB36±37%
Qwen3.6-35B-A3B-uncensored-hereticMoEQ5_K_M35.1B23.06 GiB0.02 GiB23.79 GiB0.21 GiB36±37%
Ornith-1.0-35B-uncensored-hereticMoEQ5_K_M35.1B23.06 GiB0.02 GiB23.79 GiB0.21 GiB36±37%
Nex-N2-mini-ultra-uncensored-hereticMoEQ5_K_M35.1B23.06 GiB0.02 GiB23.79 GiB0.21 GiB36±37%
Noromaid-v0.4-Mixtral-Instruct-8x7b-ZlossMoEQ3_K_L46.7B22.90 GiB0.14 GiB23.78 GiB0.22 GiB12±37%
GLM-4.7-FlashMoEQ6_K31.2B23.00 GiB0.06 GiB23.76 GiB0.24 GiB28±37%
Carnice-Qwen3.6-MoE-35B-A3BMoEI1-Q5_K_M36.0B23.03 GiB0.02 GiB23.76 GiB0.24 GiB36±37%
Qwen35B-Agent-R2-AbliteratedMoEI1-Q5_K_M34.7B23.03 GiB0.02 GiB23.76 GiB0.24 GiB36±37%
Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-DistilledMoEI1-Q5_K_M36.0B23.03 GiB0.02 GiB23.76 GiB0.24 GiB36±37%
Darwin-35B-A3B-OpusMoEI1-Q5_K_M36.0B23.03 GiB0.02 GiB23.76 GiB0.24 GiB36±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 AMD Ryzen AI Max 390 (Radeon 8050S) run?
1998 of 2118 indexed open-weight models fit a AMD Ryzen AI Max 390 (Radeon 8050S) at 4,096 context with q4_0 KV cache, the largest being NVIDIA-Nemotron-Nano-12B-v2 at BF16. That covers text, vision-language, image, video and speech models.
How much usable memory does a AMD Ryzen AI Max 390 (Radeon 8050S) actually have?
Its nameplate is 32 GB, but about 22.32 GiB is available to a model once driver and compositor overhead is accounted for, and only 24 GB of the pool can be allocated to the GPU at all.
Is a AMD Ryzen AI Max 390 (Radeon 8050S) fast for local AI?
Its memory bandwidth is 256 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.