Best Local AI Models for NVIDIA RTX 4090 (24 GB VRAM)

Q: What is the VRAM and memory bandwidth of the NVIDIA RTX 4090?

The NVIDIA RTX 4090 features 24 GB of VRAM and a memory bandwidth of 1008 GB/s (GDDR6X 384-bit). In local language model inference, VRAM determines which model sizes fit on the card, while memory bandwidth dictates how many tokens per second the GPU generates.

Q: What is the best local AI model to run on a NVIDIA RTX 4090?

The best overall model for the NVIDIA RTX 4090 is Qwen/Qwen3-8B-GGUF using the recommended Q8_0 quantization (9.47 GB total memory). With 24 GB of VRAM, this GPU typically runs a 30–35B model at Q5, fully on the GPU at full GPU speed. Check our guide to the best LLMs for 24 GB VRAM for details.

The NVIDIA RTX 4090 has 24 GB of VRAM. Here are the popular AI models it can run locally (4,096-token context, ~32.0 GB system RAM assumed), ranked by popularity.

Verdict: The flagship 24 GB gold standard: over 1 TB/s bandwidth delivers blistering 100+ tokens/s on 8B–14B models.

VRAM & Memory

24 GB

GDDR6X 384-bit

Bandwidth

1008 GB/s

Ada Lovelace

Fits in VRAM

36 models

Zero offload

TDP / Assumed RAM

450W

RAM: 32.0 GB

At a glance — 36 of the 38 most popular models run fully in VRAM on the NVIDIA RTX 4090 (24 GB), with about 168 tokens/s estimated on an 8B model in Q4. A solid pick is Qwen/Qwen3-8B-GGUF (Q8_0, 9.47 GB, ~124.3 t/s estimated).

The NVIDIA RTX 4090 is built on NVIDIA's Ada Lovelace architecture featuring GDDR6X 384-bit delivering 1008 GB/s of raw memory bandwidth. The flagship 24 GB gold standard: over 1 TB/s bandwidth delivers blistering 100+ tokens/s on 8B–14B models.

Equipped with 24 GB of dedicated VRAM, the NVIDIA RTX 4090 can run 36 popular open-source models completely in GPU memory without offloading. This includes full-speed execution for weights like Qwen3-8B-GGUF, Qwen3-4B-GGUF, Qwen2.5-7B-Instruct-GGUF. For a comprehensive breakdown of compatible model weights, see our guide to the best LLMs for 24 GB VRAM.

With a memory bandwidth of 1008 GB/s, this card can generate tokens at an estimated peak rate of ~168.0 tokens/second on an 8B parameter model (Q4_K_M). Its rated power draw is 450W TDP, so ensure your system's power supply and case ventilation are adequate for sustained local inferencing.

New to this? Read: How much VRAM do you need?

How the NVIDIA RTX 4090 compares with similar cards

Compared with the NVIDIA RTX 3090 Ti, the NVIDIA RTX 4090 has the same 24 GB of VRAM, so the same models fit fully on the GPU; the same 1008 GB/s of memory bandwidth, so the same estimated ~168 tokens/s on an 8B model in Q4 and a newer architecture (Ada Lovelace vs Ampere). For local LLMs the two generate tokens at the same speed: the difference comes down to GPU compute, which mainly affects prompt processing. Side-by-side comparison
Compared with the NVIDIA RTX 3090, the NVIDIA RTX 4090 has the same 24 GB of VRAM, so the same models fit fully on the GPU; about 8% faster token generation (~168 vs ~156 tokens/s on an 8B model in Q4, from 1008 vs 936 GB/s); a higher board power (450 vs 350 W) and a newer architecture (Ada Lovelace vs Ampere). Side-by-side comparison
Compared with the AMD Radeon RX 7900 XTX, the NVIDIA RTX 4090 has the same 24 GB of VRAM, so the same models fit fully on the GPU; about 5% faster token generation (~168 vs ~160 tokens/s on an 8B model in Q4, from 1008 vs 960 GB/s); a higher board power (450 vs 355 W) and a different architecture (Ada Lovelace vs RDNA 3). Side-by-side comparison

Token speeds here are a conservative estimate (memory bandwidth ÷ 6, for an 8B model in Q4 that reads about 5 GB of weights per token, plus overhead); prompt processing depends on GPU compute instead.

36 fit fully in VRAM · 2 run with offload

Only show models that fit fully in VRAM

Model	Size	Quant.	Quality	Memory	Speed~	Verdict
Qwen/Qwen3-8B-GGUF	8.19B	Q8_0	Excellent	9.47 GB	124.3 t/s	Fits in VRAM
Qwen/Qwen3-4B-GGUF	4.02B	Q8_0	Excellent	5.35 GB	252.9 t/s	Fits in VRAM
Qwen/Qwen2.5-7B-Instruct-GGUF	7.62B	GGUF	Excellent	15.21 GB	71.0 t/s	Fits in VRAM
unsloth/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF	8.0B	BF16	Excellent	14.87 GB	71.9 t/s	Fits in VRAM
unsloth/gpt-oss-20b-GGUF	20.91B	F16	Very good	13.74 GB	78.5 t/s	Fits in VRAM
bartowski/Qwen2.5-32B-Instruct-GGUF	32.76B	Q5_K_L	Very good	23.91 GB	45.6 t/s	Fits in VRAM
Qwen/Qwen2.5-1.5B-Instruct-GGUF	1.54B	GGUF	Excellent	4.23 GB	304.0 t/s	Fits in VRAM
unsloth/LFM2.5-24B-A2B-GGUF	23.97B	Q6_K_XL	Excellent	21.24 GB	49.5 t/s	Fits in VRAM
unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF	30.53B	Q5_K_XL	Very good	21.42 GB	49.8 t/s	Fits in VRAM
Qwen/Qwen2.5-0.5B-Instruct-GGUF	0.49B	GGUF	Excellent	2.03 GB	854.6 t/s	Fits in VRAM
unsloth/Ornith-1.0-9B-GGUF	—	BF16	Excellent	17.61 GB	60.4 t/s	Fits in VRAM
Qwen/Qwen3-14B-GGUF	14.77B	Q8_0	Excellent	16.05 GB	68.9 t/s	Fits in VRAM
Qwen/Qwen2.5-Coder-7B-Instruct-GGUF	7.62B	GGUF	Excellent	15.21 GB	71.0 t/s	Fits in VRAM
ggml-org/gpt-oss-120b-GGUF	116.83B	BF16	Very low	2.42 GB	682.1 t/s	Fits in VRAM
LiquidAI/LFM3-1.7B-GGUF	1.73B	BF16	Excellent	4.08 GB	312.2 t/s	Fits in VRAM
unsloth/flint-GGUF	39.79B	Q3_K_XL	Fair	18.13 GB	58.5 t/s	Fits in VRAM
janhq/Jan-v3.5-4B-gguf	4.41B	GGUF	Excellent	9.59 GB	122.6 t/s	Fits in VRAM
Qwen/Qwen2.5-3B-Instruct-GGUF	3.09B	GGUF	Excellent	7.27 GB	159.2 t/s	Fits in VRAM
unsloth/GLM-4.7-Flash-GGUF	31.22B	Q5_K_XL	Very good	21.21 GB	49.9 t/s	Fits in VRAM
bartowski/Meta-Llama-3.1-8B-Instruct-GGUF	8.03B	Q8_0	Excellent	9.23 GB	126.7 t/s	Fits in VRAM
unsloth/Llama-3.2-3B-Instruct-GGUF	3.21B	F16	Excellent	7.23 GB	168.2 t/s	Fits in VRAM
LiquidAI/LFM2.5-1.2B-Instruct-GGUF	1.17B	BF16	Excellent	3.03 GB	461.9 t/s	Fits in VRAM
lmstudio-community/gemma-3-1b-it-GGUF	1.0B	Q8_0	Excellent	1.96 GB	1012.2 t/s	Fits in VRAM
bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF	34.66B	Q5_K_S	Very good	23.38 GB	44.8 t/s	Fits in VRAM
ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF	31.58B	Q4_0	Good	18.42 GB	57.3 t/s	Fits in VRAM
Qwen/Qwen3-0.6B-GGUF	0.75B	Q8_0	Excellent	1.83 GB	1692.6 t/s	Fits in VRAM
bartowski/Qwen_Qwen3-0.6B-GGUF	0.75B	BF16	Excellent	2.64 GB	717.1 t/s	Fits in VRAM
bartowski/Llama-3.3-70B-Instruct-GGUF	70.55B	IQ2_S	Very low	22.93 GB	48.7 t/s	Fits in VRAM
unsloth/Qwen3-1.7B-GGUF	1.72B	BF16	Excellent	4.45 GB	314.0 t/s	Fits in VRAM
LiquidAI/LFM2.5-350M-GGUF	0.35B	BF16	Excellent	1.51 GB	1521.2 t/s	Fits in VRAM
Qwen/Qwen2.5-Coder-14B-Instruct-GGUF	14.77B	Q8_0	Excellent	16.17 GB	68.9 t/s	Fits in VRAM
unsloth/Qwen3-Coder-Next-GGUF	79.67B	IQ2_XXS	Very low	22.89 GB	46.4 t/s	Fits in VRAM
unsloth/DeepSeek-R1-0528-Qwen3-8B-GGUF	8.19B	BF16	Excellent	16.63 GB	66.0 t/s	Fits in VRAM
bartowski/Nanbeige_Nanbeige4.3-33B-GGUF	33.12B	Q5_K_S	Very good	23.65 GB	45.7 t/s	Fits in VRAM
Qwen/Qwen2.5-Coder-1.5B-Instruct-GGUF	1.54B	Q8_0	Excellent	2.67 GB	571.3 t/s	Fits in VRAM
bartowski/Qwen2.5-14B-Instruct-GGUF	14.77B	Q8_0	Excellent	16.17 GB	68.9 t/s	Fits in VRAM
unsloth/Ornith-1.5-122B-GGUF	125.09B	IQ3_XXS	Low	47.01 GB	1.1 t/s	Offload
google/gemma-7b	8.54B	GGUF	Excellent	33.11 GB	1.6 t/s	Offload

"Fits in VRAM" = fast, fully on GPU. "Offload" = part on system RAM, slower. Speed is a rough estimate from the NVIDIA RTX 4090's 1008 GB/s memory bandwidth (offloaded models are limited by system RAM).

Frequently asked questions

What is the VRAM and memory bandwidth of the NVIDIA RTX 4090?

The NVIDIA RTX 4090 features 24 GB of VRAM and a memory bandwidth of 1008 GB/s (GDDR6X 384-bit). In local language model inference, VRAM determines which model sizes fit on the card, while memory bandwidth dictates how many tokens per second the GPU generates.

What is the best local AI model to run on a NVIDIA RTX 4090?

The best overall model for the NVIDIA RTX 4090 is Qwen/Qwen3-8B-GGUF using the recommended Q8_0 quantization (9.47 GB total memory). With 24 GB of VRAM, this GPU typically runs a 30–35B model at Q5, fully on the GPU at full GPU speed. Check our guide to the best LLMs for 24 GB VRAM for details.

Can the NVIDIA RTX 4090 run 70B parameter models like Llama 3.3 70B?

Yes. With 24 GB of VRAM, the NVIDIA RTX 4090 can run 70B parameter models quantized to 3-bit or 4-bit (such as IQ3_M or Q4_K_M) with moderate context windows (4k tokens). For larger context windows (16k+), pairing with a second GPU or slight CPU offloading may be required.

What power supply (PSU) do you need for the NVIDIA RTX 4090 when running local AI?

The NVIDIA RTX 4090 has a rated TDP of 450W. While local inference typically draws less power than full 3D rasterization or gaming, continuous token generation can sustain high loads. A quality power supply of at least 750W is strongly recommended.