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. 604 of 2118 indexed models fit at 16K context with q8_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 477audio tts 19vision language 48audio asr 36embedding 22video 2

What fits at 16K context

largest quantization that fits, per model · 604 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
LFM2-2.6B-ExpQ8_02.6B2.68 GiB0.13 GiB3.72 GiB0.00 GiB28±26.5%
Dolphin3.0-Llama3.2-3BQ4_K_M3.2B1.88 GiB0.93 GiB3.72 GiB0.00 GiB28±26.5%
Llama-Doctor-3.2-3B-InstructI1-Q4_K_M3.2B1.88 GiB0.93 GiB3.72 GiB0.00 GiB28±26.5%
Llama-Song-Stream-3B-InstructQ4_K_M3.2B1.88 GiB0.93 GiB3.72 GiB0.00 GiB28±26.5%
Llama-3.2-3B-Instruct-roleplay-tunedI1-Q4_K_M3.2B1.88 GiB0.93 GiB3.72 GiB0.00 GiB28±26.5%
Llama-3.2-3B-Instruct-heretic-ablitered-uncensoredI1-Q4_K_M3.2B1.88 GiB0.93 GiB3.72 GiB0.00 GiB28±26.5%
llama-3.2-Korean-Bllossom-3BQ4_K_M3.2B1.88 GiB0.93 GiB3.72 GiB0.00 GiB28±26.5%
Llama-3.2-3B-Instruct-uncensoredQ4_K_M3.2B1.88 GiB0.93 GiB3.72 GiB0.00 GiB28±26.5%
llama-3.2-3b-instruct-bnb-4bitQ4_K_M3.3B1.88 GiB0.93 GiB3.72 GiB0.00 GiB28±26.5%
Llama-3.2-3B-InstructQ4_K_M3.2B1.88 GiB0.93 GiB3.72 GiB0.00 GiB28±26.5%
Llama-3.2-3BQ4_K_M3.2B1.88 GiB0.93 GiB3.72 GiB0.00 GiB28±26.5%
OneLLM-Doey-ChatQA-V1-Llama-3.2-3BQ4_K_M3.2B1.88 GiB0.93 GiB3.72 GiB0.00 GiB28±26.5%
Llama3.2-3B-creative-writer-v0.1I1-Q4_K_M3.2B1.88 GiB0.93 GiB3.72 GiB0.00 GiB28±26.5%
Firefly-V3.2I1-Q4_K_M3.2B1.88 GiB0.93 GiB3.72 GiB0.00 GiB28±26.5%
Firefly-V3I1-Q4_K_M3.2B1.88 GiB0.93 GiB3.72 GiB0.00 GiB28±26.5%
Hermes-3-Llama-3.2-3BQ4_K_M3.2B1.88 GiB0.93 GiB3.72 GiB0.00 GiB28±26.5%
Llama-3.2-3B-bnb-4bitQ4_K_M3.3B1.88 GiB0.93 GiB3.72 GiB0.00 GiB28±26.5%
DeepHat-V1-7B-Heretic-AbliteratedI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
ShizhenGPT-7B-VLI1-IQ2_XS8.3B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
HuatuoGPT-o1-7BI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
MathSmith-DS-Qwen-7B-LongCoTI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
AstraGPTCoder-7BI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Qwen2.5-Coder-7B-Instruct-Ghidra-v2I1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
EsDrac-v1-7BI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Hemlock-Apothecary-7B-GRPO-e3I1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
openhands-lm-7b-v0.1I1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Hemlock2-Coder-7B-GRPOI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
shellwhiz-7bI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Qwen2.5-Coder-7B-Instruct-abliteratedI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Qwen2.5-Coder-7B-Instruct-OBLITERATED-advancedI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Qwen-STEM-Specialist-7BI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
VulnLLM-R-7BI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Garnet-OCR-7B-0422I1-IQ2_XS8.3B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
UwU-7B-InstructI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Video-R1-7BI1-IQ2_XS8.3B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
HARC-Qwen2.5-7B-InstructI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Qwen2.5-Coder-7B-AbliteratedI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Bozdogan-7BI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Crazy-AI-ModelI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
turbo-ai-7bI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
DeepSeek-R1-Distill-Qwen-7B-abliterated-v2I1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Ghosty-7BI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
SP-7BI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Qwen2.5-Coder-7B-Instruct-UncensoredI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Qwen2.5-VL-7B-Instruct-abliteratedI1-IQ2_XS8.3B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
DeepSeek-R1-Distill-Qwen-8B-AbliteratedI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Qwen2.5-7BIQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Qwen2.5-VL-7B-Instruct-hereticI1-IQ2_XS8.3B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
AWARES-Qwen2.5-VL-7BI1-IQ2_XS8.3B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
olmOCR-2-7B-1025I1-IQ2_XS8.3B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Qwen2.5-7B-Instruct-UncensoredI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Med-RwRI1-IQ2_XS8.3B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
EVA-Qwen2.5-7B-v0.1I1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
SpatialThinker-7BI1-IQ2_XS8.3B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Human-Like-Qwen2.5-7B-InstructI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Qwen2-7B-InstructIQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
Hercules-5.0-Qwen2-7BIQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 GiB29±26.5%
orpheus-3b-0.1-pretrainedQ3_K_S3.8B1.88 GiB0.93 GiB3.72 GiB0.00 GiB28±26.5%
gemma-3n-E4B-itUD-IQ2_XXS7.8B2.64 GiB0.15 GiB3.72 GiB0.00 GiB29±26.5%
MQ-Coldbrew-BaseI1-IQ2_XS7.6B2.30 GiB0.46 GiB3.72 GiB0.00 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?
604 of 2118 indexed open-weight models fit a Radeon Pro W6400 at 16,384 context with q8_0 KV cache, the largest being LFM2-2.6B-Exp at Q8_0. 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.