Skip to content
Local AI Wizard
Your machine
Theme

RTX 3090 vs RTX 4090

RTX 3090 · 24 GB VRAM and RTX 4090 · 24 GB VRAM side by side for local AI.

Facts from both hardware pages

RTX 3090 vs RTX 4090: memory, bandwidth, models and prices
SpecificationRTX 3090RTX 4090
Memory24 GB VRAMDedicated · system RAM does not add · source24 GB VRAMDedicated · system RAM does not add · source
Memory bandwidth936 GB/sManufacturer spec · not model speed · source1,008 GB/sManufacturer spec · not model speed · source
Text · run well639+12 tight fit · 124 run slowly · within 24 GB639+12 tight fit · 124 run slowly · within 24 GB
Vision · run well960+33 tight fit · within 24 GB960+33 tight fit · within 24 GB
Image · within memory135within 24 GB · documented paths and minimums135within 24 GB · documented paths and minimums
Video · within memory60within 24 GB · documented paths and minimums60within 24 GB · documented paths and minimums
Audio · within memory372within 24 GB · documented paths and minimums372within 24 GB · documented paths and minimums
Embeddings · within memory650within 24 GB · 8-bit estimate650within 24 GB · 8-bit estimate
Published speed runs1413 on this machine · 1 naming only the chip1313 on this machine
Launch price (MSRP)USD 1,499US · dated 2020-09-24 · not a tracked offer · sourceUSD 1,599US · dated 2022-10-12 · not a tracked offer · source

Counts use the same memory estimate as each hardware page: text and vision run well within the fast memory (tight fits counted apart); image, video and audio use documented paths and publisher minimums. Formulas

Top text models each runs well

The first 5 rows of each hardware page's Text list, in its order: LiveBench score, then reviewed models, then Hub downloads. Memory is the planning estimate on that machine.

RTX 3090 vs RTX 4090: top text models that run well
RankRTX 3090RTX 4090
#1Smaug MiniAbacus AI · LiveBench 76.9 · ~23 GBSmaug MiniAbacus AI · LiveBench 76.9 · ~23 GB
#2Qwen3.8 27BAlibaba · LiveBench 75.3 · ~23 GBQwen3.8 27BAlibaba · LiveBench 75.3 · ~23 GB
#3Qwen3.6 27BAlibaba · LiveBench 64 · ~23 GBQwen3.6 27BAlibaba · LiveBench 64 · ~23 GB
#4Qwen3 14BAlibaba · ~14 GBQwen3 14BAlibaba · ~14 GB
#5Qwen3.5 9BAlibaba · ~9 GBQwen3.5 9BAlibaba · ~9 GB

Measured speed

Published output tokens per second, as measured. A speed is compared only when the same checkpoint, quantization, runtime and lengths ran on both machines themselves; same-chip references are not compared.

RTX 3090 vs RTX 4090: the same published test on both
Same testRTX 3090RTX 4090
gpt-oss 20B Q4_K_M · llama.cpp CUDA214.34 ± 1.512026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0277.95 ± 1.502026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
Qwen3-Coder 30B-A3B Q4_K_M · llama.cpp CUDA206.29 ± 1.402026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0265.96 ± 2.532026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
gpt-oss 20B Q8_0 · llama.cpp CUDA204.22 ± 1.072026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0258.4 ± 1.922026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
gpt-oss-20b MXFP4 · llama.cpp CUDA161.77 ± 0.56Not stated by source · llama.cpp gpt-oss guide · maintainer result221.95 ± 6.34Not stated by source · llama.cpp gpt-oss guide · maintainer result
Llama 2 7B Q4_0 · CUDA161.89 ± 0.18Not stated by source · llama.cpp CUDA scoreboard · community result188.96 ± 0.05Not stated by source · llama.cpp CUDA scoreboard · community result
Gemma 4 26B-A4B UD-Q4_K_M · llama.cpp CUDA140.26 ± 0.502026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0169.27 ± 1.612026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
Qwen3.5 9B Q4_K_M · llama.cpp CUDA122.08 ± 0.162026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0141.55 ± 0.582026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
Gemma 4 12B Q4_K_M · llama.cpp CUDA86.54 ± 0.162026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0101.95 ± 0.082026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
Qwen3.5 9B Q8_0 · llama.cpp CUDA85.66 ± 0.102026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.095.08 ± 0.132026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
Gemma 4 12B Q8_0 · llama.cpp CUDA57.24 ± 0.062026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.064.22 ± 0.092026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
Mistral Small 3.2 24B Q4_K_M · llama.cpp CUDA53.15 ± 0.012026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.061.56 ± 0.072026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
Qwen3.6 27B Q4_K_M · llama.cpp CUDA42.18 ± 0.022026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.048.38 ± 0.042026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
Gemma 4 31B Q4_K_M · llama.cpp CUDA38.78 ± 0.012026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.044.42 ± 0.042026-10-16 · Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
All published measurements · RTX 309013 on this machine · 1 naming only the chip
Published measurements on this hardware
gpt-oss 20B Q4_K_Munsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M…NVIDIA GeForce RTX 3090 · 24 GB (the card's only memory size; rented host)214.34 ± 1.51Q4_K_M
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
Qwen3-Coder 30B-A3B Q4_K_Munsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Q…NVIDIA GeForce RTX 3090 · 24 GB (the card's only memory size; rented host)206.29 ± 1.40Q4_K_M
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
gpt-oss 20B Q8_0unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.g…NVIDIA GeForce RTX 3090 · 24 GB (the card's only memory size; rented host)204.22 ± 1.07Q8_0
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
Llama 2 7B Q4_0Llama 2 7B Q4_0 (scoreboard model; the scor…Nvidia RTX 3090 · 24 GB (memory size not stated by the source; this chip is sold with one memory size) · coopmat2171.61 ± 2.69Q4_0
0 / 128 · Single llama-bench process;…
Vulkan
llama.cpp commit d05fe1d
Flash Attention enabled
llama.cpp Vulkan scoreboard · community result
2026-10-25
Llama 2 7B Q4_0Llama 2 7B Q4_0 (scoreboard model; the scor…NVIDIA GeForce RTX 3090 · 24 GB GDDR6X161.89 ± 0.18Q4_0
0 / 128 · Single llama-bench process;…
CUDA
llama.cpp commit c76b420
Flash Attention enabled
llama.cpp CUDA scoreboard · community result
2026-10-25
gpt-oss-20b MXFP4gpt-oss-20b-mxfp4.gguf (ggml-org/gpt-oss-20…NVIDIA GeForce RTX 3090 · 24 GB161.77 ± 0.56MXFP4 (native gpt-oss weights)
0 / 128 · Single llama-bench process;…
llama.cpp CUDA
llama.cpp commit a094f38 (6210)
Flash Attention enabled
llama.cpp gpt-oss guide · maintainer result
2026-10-20
Gemma 4 26B-A4B UD-Q4_K_Munsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B…NVIDIA GeForce RTX 3090 · 24 GB (the card's only memory size; rented host)140.26 ± 0.50UD-Q4_K_M
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
Qwen3.5 9B Q4_K_Munsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.g…NVIDIA GeForce RTX 3090 · 24 GB (the card's only memory size; rented host)122.08 ± 0.16Q4_K_M
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
Gemma 4 12B Q4_K_Munsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-…NVIDIA GeForce RTX 3090 · 24 GB (the card's only memory size; rented host)86.54 ± 0.16Q4_K_M
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
Qwen3.5 9B Q8_0unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.ggufNVIDIA GeForce RTX 3090 · 24 GB (the card's only memory size; rented host)85.66 ± 0.10Q8_0
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
Gemma 4 12B Q8_0unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-…NVIDIA GeForce RTX 3090 · 24 GB (the card's only memory size; rented host)57.24 ± 0.06Q8_0
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
Mistral Small 3.2 24B Q4_K_Munsloth/Mistral-Small-3.2-24B-Instruct-2506…NVIDIA GeForce RTX 3090 · 24 GB (the card's only memory size; rented host)53.15 ± 0.01Q4_K_M
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
Qwen3.6 27B Q4_K_Munsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M…NVIDIA GeForce RTX 3090 · 24 GB (the card's only memory size; rented host)42.18 ± 0.02Q4_K_M
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
Gemma 4 31B Q4_K_Munsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-…NVIDIA GeForce RTX 3090 · 24 GB (the card's only memory size; rented host)38.78 ± 0.01Q4_K_M
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
All published measurements · RTX 409013 on this machine
Published measurements on this hardware
gpt-oss 20B Q4_K_Munsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M…NVIDIA GeForce RTX 4090 · 24 GB (the card's only memory size; rented host)277.95 ± 1.50Q4_K_M
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
Qwen3-Coder 30B-A3B Q4_K_Munsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Q…NVIDIA GeForce RTX 4090 · 24 GB (the card's only memory size; rented host)265.96 ± 2.53Q4_K_M
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
gpt-oss 20B Q8_0unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q8_0.g…NVIDIA GeForce RTX 4090 · 24 GB (the card's only memory size; rented host)258.4 ± 1.92Q8_0
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
gpt-oss-20b MXFP4gpt-oss-20b-mxfp4.gguf (ggml-org/gpt-oss-20…NVIDIA GeForce RTX 4090 · 24 GB221.95 ± 6.34MXFP4 (native gpt-oss weights)
0 / 128 · Single llama-bench process;…
llama.cpp CUDA
llama.cpp commit a094f38 (6210)
Flash Attention enabled
llama.cpp gpt-oss guide · maintainer result
2026-10-20
Llama 2 7B Q4_0TheBloke/Llama-2-7B-GGUF · llama-2-7b.Q4_0.…NVIDIA GeForce RTX 4090 · 24 GB GDDR6X188.96 ± 0.05Q4_0
0 / 128 · Not reported by llama-bench
CUDA
llama.cpp commit 2241453
Flash Attention enabled
llama.cpp CUDA scoreboard · community result
2026-09-25
Gemma 4 26B-A4B UD-Q4_K_Munsloth/gemma-4-26B-A4B-it-GGUF/gemma-4-26B…NVIDIA GeForce RTX 4090 · 24 GB (the card's only memory size; rented host)169.27 ± 1.61UD-Q4_K_M
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
Qwen3.5 9B Q4_K_Munsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q4_K_M.g…NVIDIA GeForce RTX 4090 · 24 GB (the card's only memory size; rented host)141.55 ± 0.58Q4_K_M
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
Gemma 4 12B Q4_K_Munsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-…NVIDIA GeForce RTX 4090 · 24 GB (the card's only memory size; rented host)101.95 ± 0.08Q4_K_M
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
Qwen3.5 9B Q8_0unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-Q8_0.ggufNVIDIA GeForce RTX 4090 · 24 GB (the card's only memory size; rented host)95.08 ± 0.13Q8_0
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
Gemma 4 12B Q8_0unsloth/gemma-4-12b-it-GGUF/gemma-4-12b-it-…NVIDIA GeForce RTX 4090 · 24 GB (the card's only memory size; rented host)64.22 ± 0.09Q8_0
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
Mistral Small 3.2 24B Q4_K_Munsloth/Mistral-Small-3.2-24B-Instruct-2506…NVIDIA GeForce RTX 4090 · 24 GB (the card's only memory size; rented host)61.56 ± 0.07Q4_K_M
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
Qwen3.6 27B Q4_K_Munsloth/Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M…NVIDIA GeForce RTX 4090 · 24 GB (the card's only memory size; rented host)48.38 ± 0.04Q4_K_M
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25
Gemma 4 31B Q4_K_Munsloth/gemma-4-31B-it-GGUF/gemma-4-31B-it-…NVIDIA GeForce RTX 4090 · 24 GB (the card's only memory size; rented host)44.42 ± 0.04Q4_K_M
0 / 128 · Single llama-bench process
llama.cpp CUDA
llama.cpp commit 3e4dfa8f3 (b11658)
Flash Attention enabled
Data: Local LLM Advisor (localllmadvisor.com) · CC BY 4.0
2026-10-25

Verdict

Both have 24 GB VRAM. By the memory estimate, 639 text models run well on each. RTX 4090 lists 1,008 GB/s of memory bandwidth, 1.1× the 936 GB/s of RTX 3090 (a manufacturer spec, not a measured speed). In the same published test (gpt-oss 20B Q4_K_M, llama.cpp CUDA), RTX 3090 measured 214.3 and RTX 4090 278 output tokens per second. Launch prices on file: RTX 3090 USD 1,499 (2020-09-24) and RTX 4090 USD 1,599 (2022-10-12). The same number of text models run well on both; RTX 4090 measured faster in the shared test.

Limits and dates

Model counts are memory estimates, not speed measurements or guaranteed fits; verify runtime and checkpoint support. Bandwidth is the manufacturer's figure. Prices are launch or list prices with their dates, not current offers.

Page updated · Model catalog checked · LiveBench 2026-06-25