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. 494 of 2118 indexed models fit at 16K context with f16 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 376vision language 43audio asr 35audio tts 18video 2embedding 20

What fits at 16K context

largest quantization that fits, per model · 494 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
OpenClaude-1.7B-MergedQ4_11.7B1.07 GiB1.75 GiB3.72 GiB0.00 GiB28±26.5%
gemma-3-4b-it-roleplay-tuned-v1I1-Q4_K_M4.3B2.32 GiB0.48 GiB3.72 GiB0.00 GiB29±26.5%
Gemma-3-R1984-4BQ4_K_M4.3B2.32 GiB0.48 GiB3.72 GiB0.00 GiB29±26.5%
Gemma-3-4B-VL-it-Gemini-Pro-Heretic-Uncensored-ThinkingQ4_K_M4.3B2.32 GiB0.48 GiB3.72 GiB0.00 GiB29±26.5%
gemma-3-4b-it-roleplay-tuned-v2I1-Q4_K_M4.3B2.32 GiB0.48 GiB3.72 GiB0.00 GiB29±26.5%
medgemma-1.5-4b-itQ4_K_M4.3B2.32 GiB0.48 GiB3.72 GiB0.00 GiB29±26.5%
gemma-3-4b-it-heretic-uncensored-abliterated-ExtremeI1-Q4_K_M4.3B2.32 GiB0.48 GiB3.72 GiB0.00 GiB29±26.5%
medgemma-4b-itQ4_K_M4.3B2.32 GiB0.48 GiB3.72 GiB0.00 GiB29±26.5%
gemma-3-4b-it-abliteratedQ4_K_M4.3B2.32 GiB0.48 GiB3.72 GiB0.00 GiB29±26.5%
amoral-gemma3-4B-v1Q4_K_M4.3B2.32 GiB0.48 GiB3.72 GiB0.00 GiB29±26.5%
Gemma3-4B-CodeCenturionI1-Q4_K_M4.3B2.32 GiB0.48 GiB3.72 GiB0.00 GiB29±26.5%
ArrowMint-Gemma3-4B-YUKI-v0.1I1-Q4_K_M4.3B2.32 GiB0.48 GiB3.72 GiB0.00 GiB29±26.5%
gemma-3-4b-itQ4_K_M4.3B2.32 GiB0.48 GiB3.72 GiB0.00 GiB29±26.5%
Crow-9B-HERETIC-4.6I1-IQ1_S9.4B2.28 GiB0.50 GiB3.71 GiB0.01 GiB29±26.5%
Qwen3.5-9B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKINGI1-IQ1_S9.4B2.28 GiB0.50 GiB3.71 GiB0.01 GiB29±26.5%
Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSOREDI1-IQ1_S9.4B2.28 GiB0.50 GiB3.71 GiB0.01 GiB29±26.5%
Qwen3.5-9B-Claude-4.6-OS-HERETIC-UNCENSORED-INSTRUCTI1-IQ1_S9.4B2.28 GiB0.50 GiB3.71 GiB0.01 GiB29±26.5%
Qwen3.5-9B-Claude-4.6-HighIQ-THINKING-HERETIC-UNCENSOREDI1-IQ1_S9.4B2.28 GiB0.50 GiB3.71 GiB0.01 GiB29±26.5%
NaNovel-9BI1-IQ1_S9.7B2.28 GiB0.50 GiB3.71 GiB0.01 GiB29±26.5%
Qwen3.5-9B-Unredacted-MAXI1-IQ1_S9.4B2.28 GiB0.50 GiB3.71 GiB0.01 GiB29±26.5%
Qwen3.5-9B-abliteratedI1-IQ1_S9.4B2.28 GiB0.50 GiB3.71 GiB0.01 GiB29±26.5%
Ken3.5-9BI1-IQ1_S9.7B2.28 GiB0.50 GiB3.71 GiB0.01 GiB29±26.5%
Qwen3.5-9B-gemini-3.1-opus-4.6-reasoningI1-IQ1_S9.4B2.28 GiB0.50 GiB3.71 GiB0.01 GiB29±26.5%
starcoder2-3bKV unresolvedQ6_K3.0B2.32 GiB0.47 GiB3.71 GiB0.01 GiB29±26.5%
AMD-OLMo-1B-SFT-DPOQ5_K_L1.2B0.81 GiB2.00 GiB3.71 GiB0.01 GiB28±26.5%
Dolphin3.0-Llama3.2-1BF161.2B2.31 GiB0.50 GiB3.71 GiB0.01 GiB28±26.5%
Llama-3.2-1B-Instruct-hereticF161.2B2.31 GiB0.50 GiB3.71 GiB0.01 GiB28±26.5%
Llama-3.2-1BF161.2B2.31 GiB0.50 GiB3.71 GiB0.01 GiB28±26.5%
OneLLM-Doey-ChatQA-V1-Llama-3.2-1BF161.2B2.31 GiB0.50 GiB3.71 GiB0.01 GiB28±26.5%
Llama-3.2-1B-InstructF161.2B2.31 GiB0.50 GiB3.71 GiB0.01 GiB28±26.5%
Llama-3.2-1B-Instruct-UncensoredF161.2B2.31 GiB0.50 GiB3.71 GiB0.01 GiB28±26.5%
granite-speech-4.1-2b-plusQ5_K_M2.1B1.58 GiB1.25 GiB3.71 GiB0.01 GiB28±26.5%
Qwen2-1.5BQ6_K1.5B2.37 GiB0.44 GiB3.71 GiB0.01 GiB28±26.5%
Supertron2-Reranker-2BI1-Q4_12.1B1.06 GiB1.75 GiB3.71 GiB0.01 GiB28±26.5%
Uni-MuMER-Qwen3-VL-2BI1-Q4_12.1B1.06 GiB1.75 GiB3.71 GiB0.01 GiB28±26.5%
Qwen3-VL-2B-ThinkingQ4_12.1B1.06 GiB1.75 GiB3.71 GiB0.01 GiB28±26.5%
Qwen3-VL-Reranker-2BI1-Q4_12.1B1.06 GiB1.75 GiB3.71 GiB0.01 GiB28±26.5%
Qwen3-VL-2B-InstructQ4_12.1B1.06 GiB1.75 GiB3.71 GiB0.01 GiB28±26.5%
OpenCaption-2B-VL-SFT-v1.0I1-Q4_12.1B1.06 GiB1.75 GiB3.71 GiB0.01 GiB28±26.5%
Atomight-V2.5-1.7BI1-Q4_11.7B1.06 GiB1.75 GiB3.71 GiB0.01 GiB28±26.5%
gaon-1.7b-v2-translateI1-Q4_11.7B1.06 GiB1.75 GiB3.71 GiB0.01 GiB28±26.5%
gaon-1.7b-v2-instructI1-Q4_11.7B1.06 GiB1.75 GiB3.71 GiB0.01 GiB28±26.5%
Lightning-1.7BQ4_11.7B1.06 GiB1.75 GiB3.71 GiB0.01 GiB28±26.5%
DorsetHeatwaveLLM2I1-Q4_11.7B1.06 GiB1.75 GiB3.71 GiB0.01 GiB28±26.5%
Vikhr-Gemma-2B-instructQ5_K_S2.6B1.75 GiB1.04 GiB3.70 GiB0.02 GiB29±26.5%
Gemmasutra-Mini-2B-v1I1-Q5_K_S2.6B1.75 GiB1.04 GiB3.70 GiB0.02 GiB29±26.5%
gemma-2-2b-it-abliteratedQ5_K_S2.6B1.75 GiB1.04 GiB3.70 GiB0.02 GiB29±26.5%
gemma-2-2b-itQ5_K_S2.6B1.75 GiB1.04 GiB3.70 GiB0.02 GiB29±26.5%
granite-4.0-7B-A1B-Creative-v0.1MoEI1-IQ3_S6.7B2.71 GiB0.13 GiB3.70 GiB0.02 GiB78±37%
LFM2-2.6BQ8_02.6B2.55 GiB0.25 GiB3.70 GiB0.02 GiB29±26.5%
LFM2-2.6B-TranscriptQ8_02.6B2.55 GiB0.25 GiB3.70 GiB0.02 GiB29±26.5%
LFM2-VL-3BQ8_03.0B2.55 GiB0.25 GiB3.70 GiB0.02 GiB29±26.5%
gemma-2bQ8_02.5B2.49 GiB0.28 GiB3.70 GiB0.02 GiB29±26.5%
Qwen3-1.7BQ3_K_L2.0B1.06 GiB1.75 GiB3.70 GiB0.02 GiB28±26.5%
dolphin-2_6-phi-2Q8_02.8B2.75 GiB0.00 GiB3.70 GiB0.02 GiB29±26.5%
Qwen3.5-4BQ3_K_S4.7B2.29 GiB0.50 GiB3.70 GiB0.02 GiB29±26.5%
Qwen2.5-3BQ5_03.1B2.22 GiB0.56 GiB3.69 GiB0.03 GiB29±26.5%
GRM-Kerlin-3bI1-Q5_K_S3.4B2.22 GiB0.56 GiB3.69 GiB0.03 GiB29±26.5%
Qwen2.5-Coder-3B-InstructQ5_03.1B2.22 GiB0.56 GiB3.69 GiB0.03 GiB29±26.5%
Qwen2.5-3B-InstructQ5_03.1B2.22 GiB0.56 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?
494 of 2118 indexed open-weight models fit a Radeon Pro W6400 at 16,384 context with f16 KV cache, the largest being OpenClaude-1.7B-Merged at Q4_1. 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.