Can an NVIDIA GeForce RTX 4090, 24GB run local LLMs?

Yes — 28 of the 49 current models on this site fit in its 23 GB of usable memory, and three more if you keep the window shorter than 32k. The strongest it holds, Qwen3.8 27B, runs at 44 tok/s with a 32k window, worked out from memory bandwidth. Whether that saves you money is a different question: at its $1,599 launch price and 500k tokens a day, it pays back in 104 years. Its price here is the card alone, so every figure below leaves out the PC you need to put it in. Every card on this site is set against the others here.

Run the numbers on this machine

What the GeForce RTX 4090, 24GB runs

They do not all hold the same window. The weights are a fixed size, but the key-value cache grows with every token you keep, so what is left of 23 GB after the weights is how far the context goes. On 5 of the 12 below, this machine's memory is what runs out first: Qwen3.6 35B-A3B stops soonest, at 32k. The rest reach 256k, where the calculator's list ends, or the longest setting their own context limit allows.

Speed and memory are at 32k context, the setting the calculator starts on; the memory column is the weights plus the cache for that much of it. Each speed says how it was arrived at, and every one here is estimated from this machine's 1008 GB/s of memory bandwidth rather than taken from a published benchmark. There is a page on how that sum works, and what each size needs, and one on what 24 GB runs, machine by machine. The longest context is the longest setting the calculator offers that this machine still holds the model at, cache included, and each one opens the calculator on that model at that length. A figure tagged memory is one this machine ran out of room for, and the rest are stopped by the model's own limit or by the end of the list.

For scale, the calculator starts from 80 tok/s for a hosted API and times a local machine against it. 9 of the 12 above reach it, and the slowest is 43 tok/s.

16 more fit. Ranked below the twelve above: Granite 4.2 8B, Qwen3-Coder 30B-A3B, gpt-oss-20b, Gemma 4 E4B, LFM2.5 2.6B, Devstral Small 2 24B, Ministral 3 14B, Ministral 3 8B. The other 8 have no intelligence-index score, so the ranking cannot place them: Laguna XS 2.1, LFM3 1.7B, Ornith 1.5 9B, Ornith 1.5 35B-A3B, Spark-X2.5 4B, Spark-X2.6 4B, KAT-Coder V2.5 Dev 35B-A3B, LFM2.5 24B-A2B. Their pages show what each one needs and what runs it.

The 49 counted above are the current models. This site prices 16 superseded ones as well, because people still run them, and this machine holds eight of those at 32k, strongest first: Qwen3 30B-A3B Instruct 2507, Llama 3.1 8B Instruct at Q4_K_M, Llama 3.1 8B Instruct at Q8_0, Mistral Small 3.2 24B Instruct, Qwen3 14B, Qwen3 8B, Gemma 3 27B it, Gemma 3 12B it. Each of those pages asks whether the model is still worth running.

What has been measured on the GeForce RTX 4090, 24GB

Not one of the speeds above was measured on this machine, and that is not because nobody has run a model on one. This site holds four published speeds taken on this machine, three of them on superseded models the table above leaves out. They are below, as published and again at 32k of context, which is what every other speed on this page is taken at.

A published figure is what its source reports, at the window the model was run at. The 32k figure is that measurement read at 32k of context: every token generated reads the weights and the whole key-value cache, so the longer the window the slower it runs. Where this machine does not hold a model at 32k, the row gives the longest window it does hold. The figures come from ggml-org/llama.cpp, hardware-corner.net.

Three more, at a shorter window

The table above counts what fits at 32k, the context the calculator starts on. Three more models fit this machine at a shorter one. Same weights, smaller cache.

Each window is the longest setting the calculator offers that this machine still holds the model at, and it opens the calculator on that model at that length. The memory figures are the weights plus the key-value cache at that window, against the 23 GB this machine's GPU can use.

Other machines to weigh against the GeForce RTX 4090, 24GB

Every row uses the same defaults as the figures above: 500k tokens a day at 15:1 input to output, 32k context, and each machine's strongest model that fits, counted against the same 49 current models. Graphics cards are priced as the card alone, so add the PC around one before comparing it with a complete computer.

Head to head with another card: RTX PRO 6000 Blackwell, 96GB · GeForce RTX 5090, 32GB · GeForce RTX 3090, 24GB · Radeon AI PRO R9700, 32GB · GeForce RTX 4080, 16GB · GeForce RTX 3060, 12GB. There are 111 machine match-ups on the site.

The specifics

Chip

GeForce RTX 4090 — Ada Lovelace · 24 GB GDDR6X

Memory bandwidth

1008 GB/s — 21 Gbps × 384-bit bus ÷ 8 = 1,008 GB/s (21 Gbps from TechPowerUp, 384-bit from NVIDIA).

Usable by the GPU

23 GB — Usable memory is the VRAM less 1 GB: the margin llama.cpp's automatic fitting (--fit) leaves free on each device by default (common/common.h, fit_params_target = 1 GiB). Treated as nominal GB, like the other machines here.

Power under load

450 W (published) — 450 W is the maker's rated board power for the card alone. The rest of the PC adds more, and LLM decoding usually draws less than the full rating. NVIDIA lists 315 W average gaming power. A whole Intel 265K + RTX 4090 PC measured 464-509 W at peak across llama.cpp runs, prompt processing included (Jeff Geerling).

Availability

Launch price. Superseded by the RTX 50 series, so usually bought used for less: enter what you paid.

Sources

NVIDIA, NVIDIA Newsroom, TechPowerUp, geerlingguy/ai-benchmarks, ggml-org/llama.cpp