The AI Bench
Local AI checker

What AI can your computer run?

Pick your graphics card or Mac. See which models fit in its memory, how fast they will answer, and the one command to start them.

  • 52 model families
  • 90+ cards and chips
  • 0 sign-ups
1Your machine

Pick your graphics card or chip below.

System memory (RAM)
2What for

Everyday questions, writing, summaries.

3Conversation length

8K tokens ≈ 6,100 words — a normal chat.

Your RTX 4070 (GDDR6X) with 12 GB runs 13 of the 25 chat models we track, and 7 more slowly by also using system RAM.

Best for chat & writing on RTX 4070 (GDDR6X)

Gemma 4 12B

Q4_K_M · 7.1 GB file · 12B

Runs great
RTX 4070 (GDDR6X) · 12 GB8.5 of 12.0 GB usable
  • Model (incl. 0.2 GB image reader) 7.3 GB
  • Conversation memory 0.2 GB
  • App overhead 1.0 GB
Speed (estimate)
40–65tok/sFast
Longest conversation that fits
136K tokens
Spare memory
3.5 GB
ollama run gemma4:12b

Speed is an estimate from memory bandwidth, shown as a range. How we estimate

Ranked for chat & writing

Need more memory (5)smallest first

Popular machines

All hardware →
12 GB · $460–$660 new, ~$390 refurbishedNVIDIA RTX 3060 12 GBBest for chat: Gemma 4 12B
Best for coding: Qwen3.5 4B16 GB · about $800 newNVIDIA RTX 5060 Ti 16 GBBest for chat: Gemma 4 12B
Best for coding: Qwen3.5 4B24 GB · $1,000–$1,500 usedNVIDIA RTX 3090 (used)Best for chat: Qwen3.8 27B
Best for coding: Qwen3.8 27B24 GB · $2,200–$2,800 (mostly used)NVIDIA RTX 4090Best for chat: Qwen3.8 27B
Best for coding: Qwen3.8 27B24 GB · $1,699 (24 GB / 512 GB)Mac mini (M5 Pro, 24 GB)Best for chat: Gemma 4 12B
Best for coding: Gemma 4 12B64 GB · $3,799 (64 GB / 1 TB)Mac Studio (M5 Max, 64 GB)Best for chat: Qwen3.8 27B
Best for coding: Qwen3.8 27B

How the checker works

Full method →
  1. 01

    Memory

    The model file, plus memory for the conversation, plus about 1 GB for the app — compared with what your machine can use.

  2. 02

    Speed

    Every word means reading the model from memory once. We estimate speed from memory bandwidth and check it against real measurements.

  3. 03

    Ranking

    Models that fit are ranked by independent quality scores, then pushed down if they would be slow or only just fit.

Guides

All guides →
  1. 01How much memory do you need to run AI at home?Memory decides which AI models your computer can run. Here is what uses it, and what each common size gets you.
  2. 02Ollama vs LM Studio vs Jan: which app should you use?All three run the same open models on your own computer. The difference is how you use them.
  3. 03Mac or PC for local AI? Capacity vs speed, explainedA Mac holds bigger models. A PC graphics card runs smaller models faster. Which matters more depends on what you want to run.