Hardware guide / 16 GB VRAM

What LLM can a 16 GB RTX 5080 run?

Q: What LLM can an RTX 5080 run?

With a 10 percent usable-memory reserve and 20 percent inference headroom, the generic one-card weight floor reaches a 20B INT4, 8B INT8, or 3B FP16 tier. Named models still need exact checkpoint, context, KV-cache, and runtime checks.

Q: Can an RTX 5080 run Qwen3 14B?

The one-sequence 32K INT4 profile on this page reaches 14.27 GiB, leaving only about 0.13 GiB below the 14.4 GiB usable-memory budget. That is a paper fit, not comfortable runtime margin.

Q: Can an RTX 5080 run gpt-oss-20b?

The short 4K planning profile reaches 14.29 GiB and clears this budget by about 0.11 GiB, which is too narrow to promise. The conservative 128K profile reaches 22.56 GiB and does not fit fully on the card.

Q: Can an RTX 5080 run Qwen3 32B?

Not under this fully GPU-resident 32K INT4 profile. The planning target reaches 27.93 GiB, well above the 14.4 GiB usable-memory allowance.

A 16 GB card reaches the generic 20B INT4 weight tier. Exact Qwen3 14B and short-context gpt-oss-20b profiles land so close to the budget that runtime allocation decides the result.

Open the 16 GB model finder →

Fast answer

The generic weight-only floor

INT4
20B listed tier · 25.77B arithmetic ceiling

INT8
8B listed tier · 12.88B arithmetic ceiling

FP16 / BF16
3B listed tier · 6.44B arithmetic ceiling

Default reserve
16 × 0.90 ÷ 1.20 = 12 GiB for weights

Editable evidence

The 16 GB tier has two razor-thin paper fits.

Every model card opens an editable architecture-aware calculator rather than hiding the assumptions.

Qwen3 8B · INT4 · 32K9.98 GiBFits with useful planning margin → Qwen3 14B · INT4 · 32K14.27 GiBPaper fit with only 0.13 GiB margin → gpt-oss-20b · MXFP4 · 4K14.29 GiBPaper fit with only 0.11 GiB margin → Qwen3 8B · INT8 · 32K14.56 GiBMisses the default budget by 0.16 GiB → Qwen3 32B · INT4 · 32K27.93 GiBRequires a larger memory tier or offload → RTX 3060 12GB comparison10.8 GiBSee the lower-cost 12 GB boundary →

Decision boundary

A calculated fit can still fail allocation.

Two named profiles sit within 0.13 GiB of the default budget, so graph capture, allocator fragmentation, or runtime workspace can reverse the result.

Profile	Planning target	16 GB result
Qwen3 8B INT4 · 32K	9.98 GiB	Fits with 4.42 GiB margin
Qwen3 14B INT4 · 32K	14.27 GiB	Paper fit; only 0.13 GiB margin
gpt-oss-20b · 4K	14.29 GiB	Paper fit; only 0.11 GiB margin
Qwen3 8B INT8 · 32K	14.56 GiB	Misses by 0.16 GiB

Capacity is only gate one

0.11 GiB is not operational headroom.
Treat the gpt-oss-20b 4K result as a paper boundary, not a promise.

Context consumes VRAM.
The conservative gpt-oss 128K profile rises to 22.56 GiB.

Fit is not throughput.
Bandwidth, kernels, offload, and batching decide delivered speed.

Primary sources

Inputs you can inspect.

NVIDIA specifies 16 GB of GDDR7 memory for the RTX 5080. Model profiles reuse the primary-source architecture and checkpoint inputs documented in the Qwen3 and gpt-oss worksheets.

Official NVIDIA RTX 5080 specifications → Open the Qwen3 GPU worksheet → Open the gpt-oss hardware worksheet →

What LLM can an RTX 5080 run?

At the default reserve, the generic one-card floor reaches a 20B INT4, 8B INT8, or 3B FP16 tier. Named models still need exact checkpoint, context, cache, and runtime checks.

Can an RTX 5080 run Qwen3 14B?

The 32K INT4 planning profile reaches 14.27 GiB, about 0.13 GiB below the 14.4 GiB budget. Treat that as a paper fit requiring runtime validation.

Can an RTX 5080 run gpt-oss-20b?

The 4K profile reaches 14.29 GiB and clears the budget by only 0.11 GiB. The conservative 128K profile reaches 22.56 GiB and does not fit fully on the card.

Can an RTX 5080 run Qwen3 32B?

Not under this fully GPU-resident 32K INT4 profile. Its 27.93 GiB planning target exceeds the one-card allowance.