selfhost.directory

The index of self-hostable projects — ranked, categorized and served with live updates.
Own your data and stop giving it to big tech.

internlm logo

Can the RTX 3090 run InternLumina-U2?

InternLumina-U2·16B·GeForce RTX 3090·24 GB VRAM

Saved in this browser, used on every model page.
Yes

Yes. The RTX 3090 runs InternLumina-U2 fully in VRAM at Q8_0 (about 19 GB of 24 GB), at roughly 30 tokens/s.

best quant
Q8_0
needs
19 GB
usable
24 GB
est. speed
~30 tok/s

Estimates for an 8K context; offload assumes 32 GB of system RAM.Try other contexts on the InternLumina-U2 page →

Every quantisation of InternLumina-U2 on the RTX 3090

QuantQualityMemoryFitSpeed
Q2_K Heavy loss7.1 GBfits~98 tok/s
Q3_K_M Noticeable loss9.6 GBfits~66 tok/s
Q4_K_M Small loss11 GBfits~53 tok/s
Q5_K_M Near-lossless13 GBfits~45 tok/s
Q6_K Near-lossless15 GBfits~39 tok/s
Q8_0 pickLossless19 GBfits~30 tok/s
FP16 / BF16 Lossless33 GBoffload~3.2 tok/s

Speed is a bandwidth-based estimate for generating text (936 GB/s); prompt processing and long contexts are slower.

Measured speed on the RTX 3090

No community reports for this pair yet.

Report your own speed →

Omni models that fit the RTX 3090 comfortably

All that fit →
ModelParamsBest quantMemorySpeedLicence
12BQ8_0≈16 GBFits in memory~41 tok/sCommercial OK
ByteDance-Seed logoBAGEL-7B-MoTByteDance-Seed
14.7BQ8_0≈17 GBFits in memory~33 tok/sCommercial OK
17.5BQ8_0≈20 GBFits in memory~28 tok/sCommercial OK
10.7BQ8_0≈8.9 GBFits in memory~45 tok/sCommercial OK
12BQ8_0≈16 GBFits in memory~41 tok/sCommercial OK
9.4BQ8_0≈10 GBFits in memory~52 tok/sCommercial OK

Other GPUs that run InternLumina-U2

Frequently asked questions

Can the RTX 3090 run InternLumina-U2?
Yes. The RTX 3090 runs InternLumina-U2 fully in VRAM at Q8_0 (about 19 GB of 24 GB), at roughly 30 tokens/s.
How fast does InternLumina-U2 run on the RTX 3090?
An estimate from memory bandwidth gives about 30 tokens/s at Q8_0.
Which quantisation of InternLumina-U2 should I use on the RTX 3090?
Q8_0: the largest one that fits. Expected quality: lossless compared with the full-precision weights.

Memory and speed are estimates from the model's size and the card's memory bandwidth; real numbers depend on the runtime, drivers and context length. Community figures are medians of user reports.