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. 475 of 2118 indexed models fit at 64K 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
vision language 42text 362audio asr 32embedding 20audio tts 17video 2

What fits at 64K context

largest quantization that fits, per model · 475 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3.5-4BIQ3_XS4.7B2.24 GiB0.56 GiB3.71 GiB0.01 GiB29±26.5%
DeepHat-V1-7B-Heretic-AbliteratedI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
ShizhenGPT-7B-VLI1-IQ1_S8.3B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
HuatuoGPT-o1-7BI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
MathSmith-DS-Qwen-7B-LongCoTI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
AstraGPTCoder-7BI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Qwen2.5-Coder-7B-Instruct-Ghidra-v2I1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
EsDrac-v1-7BI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Hemlock-Apothecary-7B-GRPO-e3I1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
openhands-lm-7b-v0.1I1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Hemlock2-Coder-7B-GRPOI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
shellwhiz-7bI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Qwen2.5-Coder-7B-Instruct-abliteratedI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Qwen2.5-Coder-7B-Instruct-OBLITERATED-advancedI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Qwen-STEM-Specialist-7BI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
VulnLLM-R-7BI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Garnet-OCR-7B-0422I1-IQ1_S8.3B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
UwU-7B-InstructI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Video-R1-7BI1-IQ1_S8.3B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
HARC-Qwen2.5-7B-InstructI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Qwen2.5-Coder-7B-AbliteratedI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Bozdogan-7BI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Crazy-AI-ModelI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
turbo-ai-7bI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
DeepSeek-R1-Distill-Qwen-7B-abliterated-v2I1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Ghosty-7BI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
SP-7BI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Qwen2.5-Coder-7B-Instruct-UncensoredI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Qwen2.5-VL-7B-Instruct-abliteratedI1-IQ1_S8.3B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
DeepSeek-R1-Distill-Qwen-8B-AbliteratedI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Qwen2.5-7BIQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Qwen2.5-VL-7B-Instruct-hereticI1-IQ1_S8.3B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
AWARES-Qwen2.5-VL-7BI1-IQ1_S8.3B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
olmOCR-2-7B-1025I1-IQ1_S8.3B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Qwen2.5-7B-Instruct-UncensoredI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Med-RwRI1-IQ1_S8.3B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
EVA-Qwen2.5-7B-v0.1I1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
SpatialThinker-7BI1-IQ1_S8.3B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Human-Like-Qwen2.5-7B-InstructI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Qwen2-7B-InstructIQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
alduin-4b-it-baseI1-Q4_14.3B2.40 GiB0.40 GiB3.71 GiB0.01 GiB29±26.5%
MQ-Coldbrew-BaseI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Kepler-Reasoning-7BI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
DeepSeek-R1-Distill-Qwen-7B-Uncensored-ReasonerI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
DeepSeek-R1-Distill-Qwen-7B-UncensoredI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
DeepSeek-R1-STEM-Coder-7BI1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Kepler-8B-Instruct-v2I1-IQ1_S7.6B1.77 GiB0.98 GiB3.71 GiB0.01 GiB29±26.5%
Parable-Granite-4.1-3B-Claude-Fable-5I1-IQ3_XS3.4B1.40 GiB1.41 GiB3.71 GiB0.01 GiB28±26.5%
granite-4.0-microIQ3_XS3.4B1.40 GiB1.41 GiB3.71 GiB0.01 GiB28±26.5%
gemma-3-4b-it-roleplay-tuned-v1I1-Q4_14.3B2.39 GiB0.40 GiB3.70 GiB0.02 GiB29±26.5%
gemma-3-4b-it-roleplay-tuned-v2I1-Q4_14.3B2.39 GiB0.40 GiB3.70 GiB0.02 GiB29±26.5%
medgemma-1.5-4b-itQ4_14.3B2.39 GiB0.40 GiB3.70 GiB0.02 GiB29±26.5%
gemma-3-4b-it-heretic-uncensored-abliterated-ExtremeI1-Q4_14.3B2.39 GiB0.40 GiB3.70 GiB0.02 GiB29±26.5%
medgemma-4b-itQ4_14.3B2.39 GiB0.40 GiB3.70 GiB0.02 GiB29±26.5%
amoral-gemma3-4B-v1Q4_14.3B2.39 GiB0.40 GiB3.70 GiB0.02 GiB29±26.5%
gemma-3-4b-it-abliteratedQ4_14.3B2.39 GiB0.40 GiB3.70 GiB0.02 GiB29±26.5%
Gemma3-4B-CodeCenturionI1-Q4_14.3B2.39 GiB0.40 GiB3.70 GiB0.02 GiB29±26.5%
ArrowMint-Gemma3-4B-YUKI-v0.1I1-Q4_14.3B2.39 GiB0.40 GiB3.70 GiB0.02 GiB29±26.5%
gemma-3-4b-itQ4_14.3B2.39 GiB0.40 GiB3.70 GiB0.02 GiB29±26.5%
Gemma-3-4b-it-Uncensored-DBL-XIQ4_XS4.7B2.31 GiB0.47 GiB3.70 GiB0.02 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?
475 of 2118 indexed open-weight models fit a Radeon Pro W6400 at 65,536 context with q4_0 KV cache, the largest being Qwen3.5-4B at IQ3_XS. 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.