NVIDIA · workstation

RTX A400

RTX A400 has 4 GB of VRAM at 96 GB/s — about 3.72 GiB usable after driver and compositor overhead. 170 of 2118 indexed models fit at 128K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
4 GB
GDDR6
Bandwidth
96 GB/s
64-bit bus
Tensor FP16
11 TF
dense
TDP
50 W
$135 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 102vision language 14audio tts 14audio asr 28video 2embedding 10

What fits at 128K context

largest quantization that fits, per model · 170 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen-3.5-2b-roleplay-tunedI1-Q4_K_M2.3B1.22 GiB1.50 GiB3.72 GiB0.00 GiB20±22%
Financial-v1-2BI1-Q4_K_M1.9B1.22 GiB1.50 GiB3.72 GiB0.00 GiB20±22%
Qwen3.5-2BQ4_K_M2.3B1.22 GiB1.50 GiB3.72 GiB0.00 GiB20±22%
LFM2.5-Audio-1.5B-JPF161.5B2.67 GiB0.00 GiB3.72 GiB0.00 GiB20±22%
umt5-xxlQ3_K_S5.7B2.66 GiB0.00 GiB3.71 GiB0.01 GiB21±22%
Surogate-3.5-2BI1-IQ3_M2.8B1.22 GiB1.50 GiB3.71 GiB0.01 GiB20±22%
Qwen2.5-Omni-7BUD-IQ2_M10.7B2.66 GiB0.00 GiB3.70 GiB0.02 GiB21±22%
Qwen3.5-2BQ4_02.3B1.21 GiB1.50 GiB3.70 GiB0.02 GiB20±22%
Qwen3.5-2B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKINGI1-Q4_12.2B1.20 GiB1.50 GiB3.70 GiB0.02 GiB20±22%
Miss-MARTHA-hot-POCKET-edition-2B-OMNII1-Q4_12.3B1.20 GiB1.50 GiB3.70 GiB0.02 GiB20±22%
Huihui-Qwen3.5-2B-abliteratedI1-Q4_12.3B1.20 GiB1.50 GiB3.70 GiB0.02 GiB20±22%
Noema-2BI1-Q4_11.9B1.20 GiB1.50 GiB3.70 GiB0.02 GiB20±22%
Qwopus3.5-2B-v3-hereticI1-Q4_12.2B1.20 GiB1.50 GiB3.70 GiB0.02 GiB20±22%
moondream2F161.9B2.64 GiB0.00 GiB3.69 GiB0.03 GiB21±22%
PaddleOCR-VL-1.6Q8_0959M0.46 GiB2.25 GiB3.69 GiB0.03 GiB20±22%
PaddleOCR-VL-1.5Q8_0959M0.46 GiB2.25 GiB3.69 GiB0.03 GiB20±22%
Qwen3.5-2B-hereticQ4_K_M2.2B1.19 GiB1.50 GiB3.68 GiB0.04 GiB20±22%
Qwen3.5-2B-BaseQ4_K_M2.3B1.19 GiB1.50 GiB3.68 GiB0.04 GiB20±22%
VoxCPM2Q8_02.3B2.63 GiB0.00 GiB3.68 GiB0.04 GiB21±22%
LFM2.5-VL-1.6BQ8_01.6B1.16 GiB1.50 GiB3.68 GiB0.04 GiB21±22%
LFM2.5-1.2B-InstructQ8_01.2B1.16 GiB1.50 GiB3.68 GiB0.04 GiB21±22%
LFM-2.5-1.2b-Instruct-roleplay-tunedQ8_01.2B1.16 GiB1.50 GiB3.68 GiB0.04 GiB21±22%
lfm2.5-1.2b-noval-agenticQ8_01.2B1.16 GiB1.50 GiB3.68 GiB0.04 GiB21±22%
LFM2.5-1.2B-ThinkingQ8_01.2B1.16 GiB1.50 GiB3.68 GiB0.04 GiB21±22%
LFM2.5-1.2B-Instruct-UncensoredQ8_01.2B1.16 GiB1.50 GiB3.68 GiB0.04 GiB21±22%
LFM2-1.2B-ExtractQ8_01.2B1.16 GiB1.50 GiB3.68 GiB0.04 GiB21±22%
LFM2-1.2B-ToolQ8_01.2B1.16 GiB1.50 GiB3.68 GiB0.04 GiB21±22%
LFM2-1.2B-RAGQ8_01.2B1.16 GiB1.50 GiB3.68 GiB0.04 GiB21±22%
LFM2.5-1.2B-JP-202606Q8_01.2B1.16 GiB1.50 GiB3.68 GiB0.04 GiB21±22%
LFM2-VL-1.6BQ8_01.6B1.16 GiB1.50 GiB3.68 GiB0.04 GiB21±22%
EXAONE-Deep-7.8BIQ2_M7.8B2.63 GiB0.00 GiB3.67 GiB0.05 GiB21±22%
EXAONE-3.5-7.8B-InstructIQ2_M7.8B2.63 GiB0.00 GiB3.67 GiB0.05 GiB21±22%
moonshine-streaming-smallQ8_0140M0.18 GiB2.50 GiB3.65 GiB0.07 GiB20±22%
Qwen2.5-Omni-3BQ6_K5.5B2.60 GiB0.00 GiB3.65 GiB0.07 GiB21±22%
Wan2.1-T2V-1.3BQ5_K_M1.4B2.63 GiB0.00 GiB3.64 GiB0.08 GiB21±22%
EXAONE-3.5-2.4B-InstructQ8_02.4B2.64 GiB0.00 GiB3.64 GiB0.08 GiB21±22%
granite-4.0-h-microIQ4_XS3.2B1.64 GiB1.00 GiB3.64 GiB0.08 GiB21±22%
Wan2.2-TI2V-5B-TurboQ3_K_L5.0B2.59 GiB0.00 GiB3.63 GiB0.09 GiB21±22%
granite-4.0-7B-A1B-Creative-v0.1MoEI1-IQ2_XXS6.7B1.66 GiB1.00 GiB3.63 GiB0.09 GiB25±37%
granite-4.0-h-3b-arI1-Q3_K_L3.4B1.61 GiB1.00 GiB3.61 GiB0.11 GiB21±22%
t5-v1_1-xxlQ4_K_S4.8B2.55 GiB0.00 GiB3.60 GiB0.12 GiB21±22%
gemma-4-31B-it-qat-q4_0-unquantized-assistantQ8_0470M0.48 GiB2.07 GiB3.55 GiB0.17 GiB21±22%
gemma-4-31B-it-assistantQ8_0470M0.48 GiB2.07 GiB3.55 GiB0.17 GiB21±22%
Qwen3.5-2B-enkoQ4_K_M2.1B1.02 GiB1.50 GiB3.52 GiB0.20 GiB22±22%
harrier-oss-v1-270mQ8_0268M0.28 GiB2.25 GiB3.50 GiB0.22 GiB22±22%
Qwen3.5-0.8B-hereticQ8_0853M1.02 GiB1.50 GiB3.50 GiB0.22 GiB22±22%
VoxCPM-0.5BQ4_K728M2.45 GiB0.00 GiB3.49 GiB0.23 GiB22±22%
Holo-3.1-0.8BQ8_01.1B1.01 GiB1.50 GiB3.49 GiB0.23 GiB22±22%
Qwen3-Coder-Next-DFlashQ8_0474M0.47 GiB2.00 GiB3.47 GiB0.25 GiB22±22%
granite-4.0-h-1bQ8_01.5B1.45 GiB1.00 GiB3.43 GiB0.29 GiB22±22%
s2-proMoEQ2_K4.6B2.40 GiB0.00 GiB3.42 GiB0.30 GiB13±37%
nemotron-3.5-asr-streaming-0.6bF32638M2.38 GiB0.00 GiB3.42 GiB0.30 GiB23±22%
Wan2.2-TI2V-5BQ3_K_M5.0B2.37 GiB0.00 GiB3.41 GiB0.31 GiB23±22%
Atomight-V2.1-0.5B-InferenceF16494M0.93 GiB1.50 GiB3.41 GiB0.31 GiB23±22%
Atomight-V2.2-UltraThink-0.5B-abliteratedF16494M0.93 GiB1.50 GiB3.41 GiB0.31 GiB23±22%
Atomight-V2.2-UltraThink-0.5B-OBLITERATEDF16494M0.93 GiB1.50 GiB3.41 GiB0.31 GiB23±22%
DigitalAhmed-V3-qwen2.5-0.5BF16494M0.93 GiB1.50 GiB3.41 GiB0.31 GiB23±22%
Qwen2.5-Coder-0.5B-Instruct-abliteratedF16494M0.93 GiB1.50 GiB3.41 GiB0.31 GiB23±22%
Qwen2.5-Coder-0.5B-InstructF16494M0.93 GiB1.50 GiB3.41 GiB0.31 GiB23±22%
Qwen2.5-0.5B-InstructF16494M0.93 GiB1.50 GiB3.41 GiB0.31 GiB23±22%
●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 RTX A400 run?
170 of 2118 indexed open-weight models fit a RTX A400 at 131,072 context with f16 KV cache, the largest being Qwen-3.5-2b-roleplay-tuned at I1-Q4_K_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a RTX A400 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 RTX A400 fast for local AI?
Its memory bandwidth is 96 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.