AMD · workstation

Radeon Pro W6400

Radeon Pro W6400 has 4 GB of VRAM at 128 GB/s — about 3.72 GiB usable after driver and compositor overhead. 779 of 2118 indexed models fit at 16K context with q4_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
4 GB
GDDR6
Bandwidth
128 GB/s
64-bit bus
Tensor FP16
—
dense
TDP
50 W
$229 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 643vision language 54audio tts 19embedding 25audio asr 36video 2

What fits at 16K context

largest quantization that fits, per model · 779 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Apertus-8B-Instruct-2509UD-IQ1_M8.1B2.19 GiB0.56 GiB3.72 GiB0.00 GiB29±26.5%
gemma-4-E2B-it-ultra-uncensored-hereticI1-Q2_K5.1B2.78 GiB0.04 GiB3.72 GiB0.00 GiB28±26.5%
Barcenas-E2BI1-Q2_K5.1B2.78 GiB0.04 GiB3.72 GiB0.00 GiB28±26.5%
gemma-4-E2B-it-Uncensored-MAXI1-Q2_K5.1B2.78 GiB0.04 GiB3.72 GiB0.00 GiB28±26.5%
Firefly-v4I1-Q2_K5.1B2.78 GiB0.04 GiB3.72 GiB0.00 GiB28±26.5%
gemma-4-E2B-it-abliteratedI1-Q2_K5.1B2.78 GiB0.04 GiB3.72 GiB0.00 GiB28±26.5%
gemma-4-E2B-it-uncensoredQ2_K5.1B2.78 GiB0.04 GiB3.72 GiB0.00 GiB28±26.5%
gemma-4-E2B-it-heretic-araQ2_K5.1B2.78 GiB0.04 GiB3.72 GiB0.00 GiB28±26.5%
rp-model_E2B_v5.1I1-Q2_K5.1B2.78 GiB0.04 GiB3.72 GiB0.00 GiB28±26.5%
BartaLens-E2BI1-Q2_K5.1B2.78 GiB0.04 GiB3.72 GiB0.00 GiB28±26.5%
gemma-4-E2BQ2_K5.1B2.78 GiB0.04 GiB3.72 GiB0.00 GiB28±26.5%
gemma-4-E2B-itIQ4_XS5.1B2.78 GiB0.04 GiB3.71 GiB0.01 GiB28±26.5%
Holo-3.1-4BI1-IQ4_XS5.2B2.66 GiB0.14 GiB3.71 GiB0.01 GiB29±26.5%
AfriqueQwen3.5-4BI1-IQ4_XS5.2B2.66 GiB0.14 GiB3.71 GiB0.01 GiB29±26.5%
TimeOmni-1-4BI1-IQ4_XS5.2B2.66 GiB0.14 GiB3.71 GiB0.01 GiB29±26.5%
stable-code-3bI1-Q3_K_L2.8B1.40 GiB1.41 GiB3.71 GiB0.01 GiB28±26.5%
rocket-3BQ3_K_L2.8B1.40 GiB1.41 GiB3.71 GiB0.01 GiB28±26.5%
Fara1.5-4BQ4_14.5B2.66 GiB0.14 GiB3.71 GiB0.01 GiB29±26.5%
AREX-TurboQ4_14.5B2.66 GiB0.14 GiB3.71 GiB0.01 GiB29±26.5%
gemma-4-E2B-it-qat-q4_0-unquantized-hereticI1-Q2_K5.1B2.78 GiB0.04 GiB3.71 GiB0.01 GiB28±26.5%
Huihui-gemma-4-E2B-it-qat-q4_0-unquantized-abliteratedI1-Q2_K5.1B2.78 GiB0.04 GiB3.71 GiB0.01 GiB28±26.5%
Gemma4_E2B_Abliterated_Baked_HF_ReadyI1-Q2_K5.1B2.78 GiB0.04 GiB3.71 GiB0.01 GiB28±26.5%
MARTHA-LXVII.8B_QWEN-3.5-3.6_prune_9b-3.6_baseI1-IQ2_XXS8.1B2.65 GiB0.12 GiB3.71 GiB0.01 GiB29±26.5%
granite-3.1-8b-instructTQ2_08.2B2.07 GiB0.70 GiB3.70 GiB0.02 GiB29±26.5%
OpenCoder-1.5B-InstructQ8_01.9B1.89 GiB0.92 GiB3.70 GiB0.02 GiB28±26.5%
orpheus-3b-0.1-pretrainedQ4_K_M3.8B2.30 GiB0.49 GiB3.70 GiB0.02 GiB29±26.5%
Nanbeige4.2-3B-hereticQ4_K_M4.2B2.40 GiB0.39 GiB3.70 GiB0.02 GiB29±26.5%
Nanbeige4.2-3BQ4_K_M4.2B2.40 GiB0.39 GiB3.70 GiB0.02 GiB29±26.5%
Qwen3.5-4BQ4_K_S4.7B2.65 GiB0.14 GiB3.70 GiB0.02 GiB29±26.5%
LFM2.5-Queen-Opus-4.7-8B-A1BMoEI1-Q2_K_S8.5B2.75 GiB0.05 GiB3.70 GiB0.02 GiB78±37%
LFM2.5-8B-A1B-KO-SFTMoEI1-Q2_K_S8.5B2.75 GiB0.05 GiB3.70 GiB0.02 GiB78±37%
LFM2.5-8B-A1B-hereticMoEI1-Q2_K_S8.5B2.75 GiB0.05 GiB3.70 GiB0.02 GiB78±37%
LFM2.5-8B-A1B-SOMPOA-heresyMoEI1-Q2_K_S8.5B2.75 GiB0.05 GiB3.70 GiB0.02 GiB78±37%
Huihui-LFM2.5-8B-A1B-abliteratedMoEI1-Q2_K_S8.5B2.75 GiB0.05 GiB3.70 GiB0.02 GiB78±37%
Supertron2.1-8B-A1BMoEI1-Q2_K_S8.5B2.75 GiB0.05 GiB3.70 GiB0.02 GiB78±37%
dolphin-2_6-phi-2Q8_02.8B2.75 GiB0.00 GiB3.70 GiB0.02 GiB29±26.5%
FrickFritz-4BI1-Q4_14.7B2.65 GiB0.14 GiB3.70 GiB0.02 GiB29±26.5%
qwen3.5-4b-agentic-coder-v4I1-Q4_14.7B2.65 GiB0.14 GiB3.70 GiB0.02 GiB29±26.5%
Myth-4BI1-Q4_14.3B2.65 GiB0.14 GiB3.70 GiB0.02 GiB29±26.5%
Qwen3.5-4B-UncensoredI1-Q4_14.7B2.65 GiB0.14 GiB3.70 GiB0.02 GiB29±26.5%
JOSIE-2-4B-PreviewI1-Q4_14.7B2.65 GiB0.14 GiB3.70 GiB0.02 GiB29±26.5%
Surogate-3.5-4BI1-Q4_15.3B2.65 GiB0.14 GiB3.70 GiB0.02 GiB29±26.5%
Newton-bot-3-VLM-mini-4BQ4_14.7B2.65 GiB0.14 GiB3.70 GiB0.02 GiB29±26.5%
NVIDIA-Nemotron-3-Nano-4B-BF16UD-IQ2_XXS4.0B2.03 GiB0.74 GiB3.70 GiB0.02 GiB29±26.5%
Aya-Medikal-V2I1-IQ1_M8.0B2.19 GiB0.56 GiB3.70 GiB0.02 GiB29±26.5%
granite-4.0-h-tinyMoEUD-IQ3_XXS6.9B2.79 GiB0.04 GiB3.70 GiB0.02 GiB91±37%
alduin-4b-it-baseI1-Q5_K_M4.3B2.64 GiB0.14 GiB3.70 GiB0.02 GiB29±26.5%
phi-2Q3_K_M2.8B1.38 GiB1.41 GiB3.69 GiB0.03 GiB29±26.5%
qwen-indic-v1I1-IQ2_XXS7.6B2.13 GiB0.63 GiB3.69 GiB0.03 GiB29±26.5%
granite-3.1-3b-a800m-instructMoEQ6_K_L3.3B2.54 GiB0.28 GiB3.69 GiB0.03 GiB48±37%
Qwen3-VL-8B-InstructUD-IQ1_S8.8B2.13 GiB0.63 GiB3.69 GiB0.03 GiB29±26.5%
gemma-3-4b-it-roleplay-tuned-v1I1-Q5_K_M4.3B2.64 GiB0.14 GiB3.69 GiB0.03 GiB29±26.5%
Gemma-3-R1984-4BQ5_K_M4.3B2.64 GiB0.14 GiB3.69 GiB0.03 GiB29±26.5%
Gemma-3-4B-VL-it-Gemini-Pro-Heretic-Uncensored-ThinkingQ5_K_M4.3B2.64 GiB0.14 GiB3.69 GiB0.03 GiB29±26.5%
gemma-3-4b-it-roleplay-tuned-v2I1-Q5_K_M4.3B2.64 GiB0.14 GiB3.69 GiB0.03 GiB29±26.5%
medgemma-1.5-4b-itQ5_K_M4.3B2.64 GiB0.14 GiB3.69 GiB0.03 GiB29±26.5%
gemma-3-4b-it-heretic-uncensored-abliterated-ExtremeI1-Q5_K_M4.3B2.64 GiB0.14 GiB3.69 GiB0.03 GiB29±26.5%
medgemma-4b-itQ5_K_M4.3B2.64 GiB0.14 GiB3.69 GiB0.03 GiB29±26.5%
gemma-3-4b-it-abliteratedQ5_K_M4.3B2.64 GiB0.14 GiB3.69 GiB0.03 GiB29±26.5%
amoral-gemma3-4B-v1Q5_K_M4.3B2.64 GiB0.14 GiB3.69 GiB0.03 GiB29±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 Pro W6400 run?
779 of 2118 indexed open-weight models fit a Radeon Pro W6400 at 16,384 context with q4_0 KV cache, the largest being Apertus-8B-Instruct-2509 at UD-IQ1_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a Radeon Pro W6400 actually have?
Its nameplate is 4 GB, but about 3.72 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Radeon Pro W6400 fast for local AI?
Its memory bandwidth is 128 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.