GPU Sold

What can a used RTX 3090 run? Tested speeds for 11 local AI models

Tested Nov 16, 2026 on a used RTX 3090 24 GB (300 W limit), llama.cpp b12023, CUDA build

In our test, every model we tried up to 35B parameters at 4-bit fit fully in the 3090's 24 GB. Mixture-of-experts models are the sweet spot: Qwen3 Coder 30B-A3B ran at 196.6 tok/s and Qwen3.6 35B-A3B at 150.8 tok/s. Dense 27–31B models ran at 36.4–40.1 tok/s, still faster than you can read.

Buy a used RTX 3090 24GB at a fair price

Typical sold price $1,650. Most sell between $1,551 and $1,749, shipping included.

From 287 completed US eBay sales · price updated Nov 19 · price history

See used RTX 3090s on eBay

We may earn a commission if you buy through these links. It does not change your price or the sold prices on this page. Disclosure

Speed for every model we tested

RTX 3090, 4-bit weights, one chat, 512-token prompt, 128-token reply
Qwen3.5 4BDense2.73.97,035174.9instant204.6
Qwen3.5 9BDense5.76.14,311117.6instant194.3
Phi-4 14BDense8.99.12,76677.5near instant190.6
gpt-oss 20BMoE12.111.95,443191.5instant161.3
Mistral Small 3.2 24BDense14.314.21,75449.8near instant193.5
Qwen3.6 27BDense16.816.51,39440.1near instant196.4
Gemma 4 26B-A4BMoE16.917.24,357135.9instant164.1
Qwen3.8 27BDense17.616.91,06138.8faster than you read193.5
Gemma 4 31BDense18.318.81,03436.4faster than you read197.0
Qwen3 Coder 30B-A3BMoE18.618.33,002196.6instant163.1
Qwen3.6 35B-A3BMoE22.121.52,000150.8instant159.3
How we tested

llama-bench -ngl 99 -fa 1 -p 512 -n 128 -r 2; Q4 weights; f16 KV cache; batch 1

Full method

Why MoE models are so fast on a 3090

Mixture-of-experts (MoE) models only use a few billion of their parameters for each token. The “A3B” in a name means about 3B active. So they generate much faster than dense models with a similar file size.

In our test, Qwen3 Coder 30B-A3B generated 196.6 tok/s while the dense Gemma 4 31B, a similar 18.3 GB file against 18.6 GB, generated 36.4 tok/s, about 5.4× slower. They also drew less power: about 162 W average board power for the mixture-of-experts models against about 196 W for the dense ones.

Can several people share one 3090?

In our test with Qwen3.6 35B-A3B, 4 people chatting at once each still got 85.8 tok/s; at 8 people, 53.7 tok/s each. The dense Qwen3.6 27B dropped to 15.8 tok/s each at 8.

Qwen3.6 35B-A3B: 512-token prompts, 128-token replies, all chats at once
Parallel chatsTotal tok/sPer chat tok/s
1142.8142.8
2240.8120.4
4343.085.8
8429.453.7
Qwen3.5 9B: 512-token prompts, 128-token replies, all chats at once
Parallel chatsTotal tok/sPer chat tok/s
1114.9114.9
2210.9105.5
4292.173.0
8376.347.0
Qwen3.6 27B: 512-token prompts, 128-token replies, all chats at once
Parallel chatsTotal tok/sPer chat tok/s
139.439.4
271.936.0
498.824.7
8126.615.8
How we tested

llama-batched-bench -npp 512 -ntg 128 -npl 1,2,4,8 -fa on (per-chat = total/parallel)

Full method

Longer conversations

In our test, Qwen3.6 27B fit 64K tokens of context with the default cache and 128K with an 8-bit cache. How much context fits in 24 GB

Bigger models

In our test, Qwen3 Coder Next 80B-A3B ran at 41.4 tok/s with part of the model in system RAM. Running models bigger than 24 GB

Is a used 3090 worth it for local AI?

In our test, 6 of the models needed more than 16 GB of VRAM, so a 16 GB card could not hold them fully: Qwen3.6 27B, Gemma 4 26B-A4B, Qwen3.8 27B, Gemma 4 31B, Qwen3 Coder 30B-A3B, Qwen3.6 35B-A3B. All of them fit in the 3090's 24 GB, which makes 24 GB the size that runs 26–35B models at 4-bit.

The used price moves daily. See the RTX 3090 sold price history.

Questions

How many tokens per second does an RTX 3090 get?

It depends on the model. In our test of generation speed with one chat: dense 27–31B models 36.4–40.1 tok/s; mixture-of-experts 20–35B models 135.9–196.6 tok/s; small 4–9B models 117.6–174.9 tok/s.

What is the largest model that fits on a 3090?

Fully in VRAM, the largest we ran was Qwen3.6 35B-A3B: a 22.1 GB file that used 21.5 GB of VRAM. With some experts kept in system RAM, Qwen3 Coder Next 80B-A3B ran at 41.4 tok/s.

How much power does a 3090 draw running an LLM?

159–205 W average board power in our runs, with the card limited to 300 W.

Can a 3090 run a 70B model?

We have not measured a dense 70B yet. The largest models we ran were mixture-of-experts with part of the experts in system RAM: gpt-oss 120B REAP (58B) at 34.7 tok/s and Qwen3 Coder Next 80B-A3B at 41.4 tok/s. A 2-bit 70B test is scheduled.

What Can an RTX 3090 Run? Tested LLM Speeds (2026)