Local LLM hardware / RTX 5090

What LLM can a 32 GB RTX 5090 run?

Q: What LLM can an RTX 5090 run?

With 32 GB of VRAM, 90 percent usable memory, and 20 percent runtime headroom, an RTX 5090 clears this site's generic weight floor for a 32B tier at INT4, 20B at INT8, or 8B at FP16. Named models still need architecture-aware context and runtime checks.

Q: Can an RTX 5090 run Qwen3 32B?

It clears this page's one-sequence 32K INT4 planning profile narrowly. The model reaches 27.93 GiB after an eight GiB BF16 KV cache and 20 percent runtime headroom, leaving about 0.87 GiB below the 28.8 GiB usable-memory budget used here.

Q: Can an RTX 5090 run gpt-oss-20b at 128K context?

The conservative full-cache profile on this site reaches 22.56 GiB using the official 12.8 GiB checkpoint, a six GiB BF16 cache allowance, and 20 percent runtime headroom. That clears the 28.8 GiB planning budget, though the actual serving engine and concurrency still need testing.

Q: Can an RTX 5090 run a 70B model entirely in VRAM?

Not at the ordinary four-bit weight floor used here. Seventy billion parameters require about 32.6 GiB for raw four-bit weights before KV cache, runtime workspace, metadata, and fragmentation.

A 32 GB card clears a generic 32B INT4 weight floor with more room than the RTX 4090. Qwen3 32B can also clear the 32K planning profile—but closely enough that runtime allocation still decides the result.

Open the 32 GB model finder →

Fast capacity answer

More VRAM changes the context boundary first.

INT4
32B listed tier · 51.5B arithmetic ceiling

INT8
20B listed tier · 25.8B arithmetic ceiling

FP16 / BF16
8B listed tier · 12.9B arithmetic ceiling

Default reserve
32 × 0.90 ÷ 1.20 = 24 GiB for weights

Named-model checks

Models near the RTX 5090 boundary.

The result is a capacity plan, not a tokens-per-second promise. Open any card to replace the default cache, context, reserve, or concurrency assumptions.

Qwen3 30B-A3B · INT4 · 32K 20.64 GiB Fits below the 28.8 GiB budget → Qwen3 32B · INT4 · 32K 27.93 GiB Fits with about 0.87 GiB margin → gpt-oss-20b · MXFP4 · 128K 22.56 GiB Fits the conservative full-cache profile → Qwen3 32B · INT8 · 32K 46.26 GiB Precision pushes it beyond one card → 70B · raw INT4 weights 32.60 GiB Already above the 32 GB label → RTX 4090 comparison 21.6 GiB See what changes with 24 GB →

24 GB versus 32 GB

The extra eight GB turns two near misses into fits.

Both columns reserve ten percent of the card and use the same model, cache, and headroom assumptions. Performance differences are outside this capacity comparison.

Profile	RTX 4090 · 21.6 GiB usable	RTX 5090 · 28.8 GiB usable
Qwen3 30B-A3B INT4 · 32K	Fits · 20.64 GiB	Fits · 20.64 GiB
Qwen3 32B INT4 · 32K	Does not fit · 27.93 GiB	Fits narrowly · 27.93 GiB
gpt-oss-20b · 128K	Does not fit · 22.56 GiB	Fits · 22.56 GiB
70B raw INT4 weights	Does not fit · 32.60 GiB	Does not fit · 32.60 GiB

Do not spend the margin twice

0.87 GiB is narrow.
A different artifact, graph capture, or runtime workspace can erase the Qwen3 32B margin.

Context repeats.
Concurrent sequences multiply the fully allocated KV-cache budget.

VRAM is not speed.
Capacity alone cannot prove useful latency, throughput, or software support.

Primary references and next checks

Verify the card, checkpoint, and runtime.

NVIDIA specifies 32 GB of GDDR7 memory for the RTX 5090. The model profiles reuse the primary-source architecture and checkpoint inputs documented in the Qwen3 and gpt-oss worksheets.

Official NVIDIA RTX 5090 specifications → Open the Qwen3 GPU worksheet → Open the gpt-oss hardware worksheet → Build a different model profile →

What LLM can an RTX 5090 run?

At the default reserve, the generic one-card floor reaches a 32B INT4, 20B INT8, or 8B FP16 tier. Named models still need exact checkpoint, context, and runtime checks.

Can an RTX 5090 run Qwen3 32B?

The 32K INT4 planning profile reaches 27.93 GiB, about 0.87 GiB below the 28.8 GiB usable-memory budget used here. Treat that as a narrow fit requiring runtime validation.

Can an RTX 5090 run gpt-oss-20b at 128K context?

The conservative full-cache profile reaches 22.56 GiB and clears the one-card budget. Concurrency, graph capture, and the selected engine can still change the result.

Can an RTX 5090 run a 70B model entirely in VRAM?

Not at the ordinary four-bit floor used here. Raw INT4 weights alone require about 32.6 GiB before cache, workspace, and metadata.