Lab

Can I run it on my machine?

Choose your graphics card, Mac or CPU. See which open models fit, how much memory they need and roughly how fast they write.

Live from OpenRouter and the Arena · checked 2 min ago

Arena leaderboards of 2026-10-28

1 · What are you running it on?

How many cards?

System RAM (a model that is too big for the card can spill into it, slowly)

2 · Precision

How compactly the weights are stored. Lower means smaller and faster, with a small loss of quality.

Smallest. Small loss in quality. What most people run.

3 · How long are your prompts?

Long context needs extra memory for the model's working notes.

Include older models

Your machine

GeForce RTX 4090

About 23 GB usable for a model, 1,008 GB/s, plus 24 GB of system RAM to spill into.

Best-rated model that runs

Gemma 4 31B Google

Arena rating 1443 · about 38 tokens a second (fast)

Biggest model that runs

Qwen3 32B Alibaba

32B · about 36 tokens a second

Qwen3.8 27B Alibaba
27B · Arena 1438 · 1M context
~42 tok/s · fastRuns well
19 GB needed: 16 GB weights + 1.7 GB context82% of 23 GB
Model card Or use it by API: what would it cost?
Gemma 4 26B A4B Google
26B total, 4B active per token · Arena 1434 · 262K context
~178 tok/s · very fastRuns well
18 GB needed: 16 GB weights + 1.0 GB context77% of 23 GB
Model card Or use it by API: what would it cost?
Qwen3.5-27B Alibaba
27B · Arena 1409 · 262K context
~42 tok/s · fastRuns well
19 GB needed: 16 GB weights + 1.7 GB context82% of 23 GB
Model card Or use it by API: what would it cost?
Qwen3 30B A3B Instruct 2507 Alibaba
30B total, 3B active per token · Arena 1384 · 262K context
~207 tok/s · very fastRuns well
20 GB needed: 18 GB weights + 1.0 GB context87% of 23 GB
Model card Or use it by API: what would it cost?
Hy-MT2-30B-A3B Tencent
30B total, 3B active per token · 8K context
~207 tok/s · very fastRuns well
20 GB needed: 18 GB weights + 1.0 GB context87% of 23 GB
Model card Or use it by API: what would it cost?
Nemotron 3.5 Lightning NVIDIA
30B total, 3B active per token · 262K context
~207 tok/s · very fastRuns well
20 GB needed: 18 GB weights + 1.0 GB context87% of 23 GB
Model card Or use it by API: what would it cost?
Nemotron 3 Nano Omni NVIDIA
30B total, 3B active per token · 256K context
~207 tok/s · very fastRuns well
20 GB needed: 18 GB weights + 1.0 GB context87% of 23 GB
Model card Or use it by API: what would it cost?
Nemotron 3 Nano 30B A3B NVIDIA
30B total, 3B active per token · 262K context
~207 tok/s · very fastRuns well
20 GB needed: 18 GB weights + 1.0 GB context87% of 23 GB
Model card Or use it by API: what would it cost?
Qwen3 VL 30B A3B Thinking Alibaba
30B total, 3B active per token · 262K context
~207 tok/s · very fastRuns well
20 GB needed: 18 GB weights + 1.0 GB context87% of 23 GB
Model card Or use it by API: what would it cost?
Qwen3 VL 30B A3B Instruct Alibaba
30B total, 3B active per token · 262K context
~207 tok/s · very fastRuns well
20 GB needed: 18 GB weights + 1.0 GB context87% of 23 GB
Model card Or use it by API: what would it cost?
Qwen3 Coder 30B A3B Instruct Alibaba
30B total, 3B active per token · 262K context
~207 tok/s · very fastRuns well
20 GB needed: 18 GB weights + 1.0 GB context87% of 23 GB
Model card Or use it by API: what would it cost?
Qwen3 30B A3B Alibaba
30B total, 3B active per token · 41K context
~207 tok/s · very fastRuns well
20 GB needed: 18 GB weights + 1.0 GB context87% of 23 GB
Model card Or use it by API: what would it cost?
Xing4.0 29B A4B XingChen
29B total, 4B active per token · 262K context
~178 tok/s · very fastRuns well
19 GB needed: 17 GB weights + 1.0 GB context85% of 23 GB
Model card Or use it by API: what would it cost?
Ternary Bonsai 2 27B PrismML
27B · 262K context
~42 tok/s · fastRuns well
19 GB needed: 16 GB weights + 1.7 GB context82% of 23 GB
Model card Or use it by API: what would it cost?
Qwen3.6 27B Alibaba
27B · 262K context
~42 tok/s · fastRuns well
19 GB needed: 16 GB weights + 1.7 GB context82% of 23 GB
Model card Or use it by API: what would it cost?
Gemma 3 27B Google
27B · 131K context
~42 tok/s · fastRuns well
19 GB needed: 16 GB weights + 1.7 GB context82% of 23 GB
Model card Or use it by API: what would it cost?
LFM2.5-24B-A2B Liquid AI
24B total, 2B active per token · 128K context
~246 tok/s · very fastRuns well
16 GB needed: 14 GB weights + 1.0 GB context72% of 23 GB
Model card Or use it by API: what would it cost?
Voxtral Small 24B 2507 Mistral AI
24B · 33K context
~47 tok/s · fastRuns well
17 GB needed: 14 GB weights + 1.6 GB context74% of 23 GB
Model card Or use it by API: what would it cost?
Mistral Small 3.2 24B Mistral AI
24B · 256K context
~47 tok/s · fastRuns well
17 GB needed: 14 GB weights + 1.6 GB context74% of 23 GB
Model card Or use it by API: what would it cost?
Mistral Small 3.1 24B Mistral AI
24B · 128K context
~47 tok/s · fastRuns well
17 GB needed: 14 GB weights + 1.6 GB context74% of 23 GB
Model card Or use it by API: what would it cost?
gpt-oss-safeguard-20b OpenAI
20B · 131K context
~55 tok/s · fastRuns well
15 GB needed: 12 GB weights + 1.5 GB context63% of 23 GB
Model card Or use it by API: what would it cost?
gpt-oss-20b OpenAI
20B total, 3.6B active per token · 131K context
~188 tok/s · very fastRuns well
14 GB needed: 12 GB weights + 1.0 GB context61% of 23 GB
Model card Or use it by API: what would it cost?
Ministral 3 14B 2512 Mistral AI
14B · 262K context
~75 tok/s · fastRuns well
11 GB needed: 8.4 GB weights + 1.3 GB context47% of 23 GB
Model card Or use it by API: what would it cost?
Qwen3 14B Alibaba
14B · 41K context
~75 tok/s · fastRuns well
11 GB needed: 8.4 GB weights + 1.3 GB context47% of 23 GB
Model card Or use it by API: what would it cost?
Llama Guard 4 12B Meta
12B · 164K context
~84 tok/s · very fastRuns well
9.4 GB needed: 7.2 GB weights + 1.2 GB context41% of 23 GB
Model card Or use it by API: what would it cost?
Gemma 3 12B Google
12B · 131K context
~84 tok/s · very fastRuns well
9.4 GB needed: 7.2 GB weights + 1.2 GB context41% of 23 GB
Model card Or use it by API: what would it cost?
Ternary Bonsai 2 9B PrismML
9B · 262K context
~105 tok/s · very fastRuns well
7.5 GB needed: 5.4 GB weights + 1.1 GB context33% of 23 GB
Model card Or use it by API: what would it cost?
Moondream 3.1 Moondream
9B total, 2B active per token · 32K context
~246 tok/s · very fastRuns well
7.4 GB needed: 5.4 GB weights + 1.0 GB context32% of 23 GB
Model card Or use it by API: what would it cost?
Qwen3.5-9B Alibaba
9B · 262K context
~105 tok/s · very fastRuns well
7.5 GB needed: 5.4 GB weights + 1.1 GB context33% of 23 GB
Model card Or use it by API: what would it cost?
Granite 4.2 8B IBM
8B · 131K context
~114 tok/s · very fastRuns well
6.8 GB needed: 4.8 GB weights + 1.0 GB context30% of 23 GB
Model card Or use it by API: what would it cost?
Ministral 3 8B 2512 Mistral AI
8B · 262K context
~114 tok/s · very fastRuns well
6.8 GB needed: 4.8 GB weights + 1.0 GB context30% of 23 GB
Model card Or use it by API: what would it cost?
Qwen3 VL 8B Instruct Alibaba
8B · 262K context
~114 tok/s · very fastRuns well
6.8 GB needed: 4.8 GB weights + 1.0 GB context30% of 23 GB
Model card Or use it by API: what would it cost?
Hy-MT2-7B Tencent
7B · 8K context
~126 tok/s · very fastRuns well
6.2 GB needed: 4.2 GB weights + 1.0 GB context27% of 23 GB
Model card Or use it by API: what would it cost?
UI-TARS 7B ByteDance
7B · 128K context
~126 tok/s · very fastRuns well
6.2 GB needed: 4.2 GB weights + 1.0 GB context27% of 23 GB
Model card Or use it by API: what would it cost?
Gemma 3 4B Google
4B · 131K context
~178 tok/s · very fastRuns well
4.4 GB needed: 2.4 GB weights + 1.0 GB context19% of 23 GB
Model card Or use it by API: what would it cost?
Ministral 3 3B 2512 Mistral AI
3B · 131K context
~207 tok/s · very fastRuns well
3.8 GB needed: 1.8 GB weights + 1.0 GB context17% of 23 GB
Model card Or use it by API: what would it cost?
LFM2.5-2.6B Liquid AI
2.6B · 128K context
~221 tok/s · very fastRuns well
3.6 GB needed: 1.6 GB weights + 1.0 GB context16% of 23 GB
Model card Or use it by API: what would it cost?
Hy-MT2-1.8B Tencent
1.8B · 8K context
~256 tok/s · very fastRuns well
3.1 GB needed: 1.1 GB weights + 1.0 GB context14% of 23 GB
Model card Or use it by API: what would it cost?
Gemma 4 31B Google
31B · Arena 1443 · 262K context
~38 tok/s · fastTight fit
21 GB needed: 19 GB weights + 1.8 GB context93% of 23 GB
Model card Or use it by API: what would it cost?
Qwen3 32B Alibaba
32B · Arena 1340 · 131K context
~36 tok/s · fastTight fit
22 GB needed: 19 GB weights + 1.8 GB context96% of 23 GB
Model card Or use it by API: what would it cost?

How these numbers are worked out

Memory = the weights (parameters × 0.6 GB per billion at 4-bit, the same estimate as the AI news page) + the context cache + about 1 GB for the runtime. A graphics card loses about 1 GB to the driver; a Mac lets the GPU use about two thirds (up to 36 GB) to three quarters of its memory by default.
Speed: every new token reads the active weights once, so tokens a second is about 60–70% of memory bandwidth divided by the bytes read. A mixture-of-experts model keeps all its weights in memory but reads only the active experts, so it is much faster than a dense model of the same size.
Splitting a model over several cards is never perfectly efficient (we count 75%). A model that spills into system RAM runs at a blend of card and RAM speed.
These are planning estimates, within about 40%. Software (llama.cpp, MLX, vLLM, Ollama), prompt length and the exact build change real speed. Models without a stated size are left out.

GeForce RTX 4090

Qwen3.8 27B Alibaba

Gemma 4 26B A4B Google

Qwen3.5-27B Alibaba

Qwen3 30B A3B Instruct 2507 Alibaba

Hy-MT2-30B-A3B Tencent

Nemotron 3.5 Lightning NVIDIA

Nemotron 3 Nano Omni NVIDIA

Nemotron 3 Nano 30B A3B NVIDIA

Qwen3 VL 30B A3B Thinking Alibaba

Qwen3 VL 30B A3B Instruct Alibaba

Qwen3 Coder 30B A3B Instruct Alibaba

Qwen3 30B A3B Alibaba

Xing4.0 29B A4B XingChen

Ternary Bonsai 2 27B PrismML

Qwen3.6 27B Alibaba

Gemma 3 27B Google

LFM2.5-24B-A2B Liquid AI

Voxtral Small 24B 2507 Mistral AI

Mistral Small 3.2 24B Mistral AI

Mistral Small 3.1 24B Mistral AI

gpt-oss-safeguard-20b OpenAI

gpt-oss-20b OpenAI

Ministral 3 14B 2512 Mistral AI

Qwen3 14B Alibaba

Llama Guard 4 12B Meta

Gemma 3 12B Google

Ternary Bonsai 2 9B PrismML

Moondream 3.1 Moondream

Qwen3.5-9B Alibaba

Granite 4.2 8B IBM

Ministral 3 8B 2512 Mistral AI

Qwen3 VL 8B Instruct Alibaba

Hy-MT2-7B Tencent

UI-TARS 7B ByteDance

Gemma 3 4B Google

Ministral 3 3B 2512 Mistral AI

LFM2.5-2.6B Liquid AI

Hy-MT2-1.8B Tencent

Gemma 4 31B Google

Qwen3 32B Alibaba

More in the Lab

GeForce RTX 4090

Qwen3.8 27B Alibaba

Gemma 4 26B A4B Google

Qwen3.5-27B Alibaba

Qwen3 30B A3B Instruct 2507 Alibaba

Hy-MT2-30B-A3B Tencent

Nemotron 3.5 Lightning NVIDIA

Nemotron 3 Nano Omni NVIDIA

Nemotron 3 Nano 30B A3B NVIDIA

Qwen3 VL 30B A3B Thinking Alibaba

Qwen3 VL 30B A3B Instruct Alibaba

Qwen3 Coder 30B A3B Instruct Alibaba

Qwen3 30B A3B Alibaba

Xing4.0 29B A4B XingChen

Ternary Bonsai 2 27B PrismML

Qwen3.6 27B Alibaba

Gemma 3 27B Google

LFM2.5-24B-A2B Liquid AI

Voxtral Small 24B 2507 Mistral AI

Mistral Small 3.2 24B Mistral AI

Mistral Small 3.1 24B Mistral AI

gpt-oss-safeguard-20b OpenAI

gpt-oss-20b OpenAI

Ministral 3 14B 2512 Mistral AI

Qwen3 14B Alibaba

Llama Guard 4 12B Meta

Gemma 3 12B Google

Ternary Bonsai 2 9B PrismML

Moondream 3.1 Moondream

Qwen3.5-9B Alibaba

Granite 4.2 8B IBM

Ministral 3 8B 2512 Mistral AI

Qwen3 VL 8B Instruct Alibaba

Hy-MT2-7B Tencent

UI-TARS 7B ByteDance

Gemma 3 4B Google

Ministral 3 3B 2512 Mistral AI

LFM2.5-2.6B Liquid AI