NVIDIA · consumer

GeForce RTX 3050

GeForce RTX 3050 has 8 GB of VRAM at 224 GB/s — about 7.44 GiB usable after driver and compositor overhead. 986 of 2118 indexed models fit at 32K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
8 GB
GDDR6
Bandwidth
224 GB/s
128-bit bus
Tensor FP16
36 TF
dense
TDP
130 W
$249 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 806embedding 25vision language 88audio tts 20video 8audio asr 38image 1

What fits at 32K context

largest quantization that fits, per model · 986 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
OLMoE-1B-7B-0924-InstructMoEI1-IQ3_XS6.9B2.67 GiB4.00 GiB7.44 GiB0.00 GiB21±37%
Qwen3-Reranker-4BIQ4_XS4.0B2.13 GiB4.50 GiB7.44 GiB0.00 GiB24±12.9%
Octen-Embedding-4BIQ4_XS4.0B2.13 GiB4.50 GiB7.44 GiB0.00 GiB24±12.9%
OmniAtlas-Qwen3-30B-A3BI1-IQ1_M31.7B6.59 GiB0.00 GiB7.44 GiB0.00 GiB25±12.9%
Qwen3-Omni-30B-A3B-CaptionerI1-IQ1_M31.7B6.59 GiB0.00 GiB7.44 GiB0.00 GiB25±12.9%
Qwen2-1.5BF161.5B5.76 GiB0.88 GiB7.44 GiB0.00 GiB24±12.9%
Yi-Coder-1.5B-ChatIQ3_XS1.5B0.65 GiB6.00 GiB7.44 GiB0.00 GiB24±12.9%
Yi-Coder-1.5BIQ3_XS1.5B0.65 GiB6.00 GiB7.44 GiB0.00 GiB24±12.9%
next-8bI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Supertron2-Reranker-8BI1-IQ1_M8.8B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
next-ocrI1-IQ1_M8.8B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Qwen3-VL-8B-GLM-4.7-Flash-Heretic-Uncensored-ThinkingI1-IQ1_M8.8B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Midas-FableAgent-8BI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Qwen3-VL-8B-Heretic-1.3.0I1-IQ1_M8.8B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Qwen3-VL-8B-Thinking-Unredacted-MAXI1-IQ1_M8.8B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Qwen3-VL-8B-Instruct-Minecraft-MT-en-zhI1-IQ1_M8.8B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Qwen-3-VL-8B-Instruct-hereticI1-IQ1_M8.8B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Poe-8B-GLM5-Opus4.6-Sonnet4.5-Kimi-Grok-Gemini-3-pro-preview-HERETICI1-IQ1_M8.8B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
ToolCUA-8BI1-IQ1_M8.8B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Huihui-Qwen3-VL-8B-Instruct-abliteratedI1-IQ1_M8.8B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Qwen3-VL-Reranker-8BI1-IQ1_M8.8B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Salience-1-9BI1-IQ1_M8.8B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Qwen3-VL-8B-Instruct-Uncensored-V2I1-IQ1_M8.8B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Maestro1-9BI1-IQ1_M8.8B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
GRaPE-2-FlashI1-IQ1_M8.8B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Jan-v2-VL-medI1-IQ1_M8.8B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Parable-Qwen3-8B-Claude-Fable-5I1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
ReasonCritic-7BI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
mythos-9b-unhinged-hereticI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Finch-8B-KTOI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Finch-8BI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
MathSmith-hc-Qwen3-8BI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
MiroThinker-v1.0-8BI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
mythos-9b-unhingedI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Ektome-Qwen3-8B-PristinelyUncensoredI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Marco-DeepResearch-8BI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
mythos-9b-mergedI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
qwen3-8b-apostateI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Josiefied-Qwen3-8B-abliterated-v1I1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
tmax-8bI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Qwen3-8B-abliteratedI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
story_generation_Qwen3_8B_RLI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
AReaL-boba-2-8B-OpenI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
DS-R1-Qwen3-8B-ArliAI-RpR-v4-SmallI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
S1-Base-8BI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Huihui-Qwen3-8B-abliterated-v2I1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Step3-VL-10B-BaseI1-IQ1_M10.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Qwen3-Reranker-8BI1-IQ1_M8.2B2.10 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Dolphin3.0-Qwen2.5-1.5BF321.5B5.76 GiB0.88 GiB7.43 GiB0.01 GiB24±12.9%
SmolLM2-1.7B-InstructQ2_K1.7B0.63 GiB6.00 GiB7.43 GiB0.01 GiB24±12.9%
Qwen3-VL-4B-Instruct-Unredacted-MAXI1-IQ4_XS4.4B2.11 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Qwen3-VL-4B-ThinkingIQ4_XS4.4B2.11 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Qwen3-VL-4B-Thinking-Unredacted-MAXI1-IQ4_XS4.4B2.11 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Zubr1.0-VL-4BI1-IQ4_XS4.4B2.11 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Huihui-Qwen3-VL-4B-Instruct-abliteratedI1-IQ4_XS4.4B2.11 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Qwen3-VL-4B-Instruct-UncensoredI1-IQ4_XS4.4B2.11 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Qwen3-VL-4B-InstructIQ4_XS4.4B2.11 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
OpenCaption-4B-VL-SFT-v1.0I1-IQ4_XS4.4B2.11 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Parable-Qwen3-4B-Claude-Fable-5I1-IQ4_XS4.0B2.11 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
Jan-v1-4BIQ4_XS4.0B2.11 GiB4.50 GiB7.43 GiB0.01 GiB24±12.9%
●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 GeForce RTX 3050 run?
986 of 2118 indexed open-weight models fit a GeForce RTX 3050 at 32,768 context with f16 KV cache, the largest being OLMoE-1B-7B-0924-Instruct at I1-IQ3_XS. That covers text, vision-language, image, video and speech models.
How much usable memory does a GeForce RTX 3050 actually have?
Its nameplate is 8 GB, but about 7.44 GiB is available to a model once driver and compositor overhead is accounted for.
Is a GeForce RTX 3050 fast for local AI?
Its memory bandwidth is 224 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.