The AI Bench
Zhipu AI (Z.ai) · MIT

GLM-5.3-Flash

Z.ai’s 320B model that reads images, with a 1M context. The main pick for 192 to 512 GB machines.

How much memory it needs

The model file, plus memory for the conversation, plus about 1 GB for the app. Longer conversations need more.

SizeVersionFileAt 8KAt 32KAt 128K
321B (18B active)UD-IQ3_XXS120 GB123 GB123 GB124 GB
UD-Q3_K_XL148 GB150 GB150 GB151 GB
UD-IQ4_XS157 GB159 GB159 GB161 GB
UD-Q4_K_XL200 GB202 GB202 GB204 GB
UD-Q6_K_XL292 GB294 GB294 GB296 GB
Q8_0341 GB343 GB344 GB345 GB

Which computers run it

The best size for each machine at an 8K-token conversation. Speeds are estimates from memory bandwidth.

ComputerBest versionFitSpeed
Laptop or PC, no graphics card, 32 GB321B UD-IQ3_XXSToo big—
RTX 4060 · 8 GB321B UD-IQ3_XXSToo big—
RTX 3060 · 12 GB321B UD-IQ3_XXSToo big—
RTX 5060 Ti · 16 GB321B UD-IQ3_XXSToo big—
RTX 4090 · 24 GB321B UD-IQ3_XXSToo big—
RTX 5090 · 32 GB321B UD-IQ3_XXSToo big—
Mac with M4 · 16 GB321B UD-IQ3_XXSToo big—
Mac with M4 Pro · 48 GB321B UD-IQ3_XXSToo big—
Mac with M5 Max · 128 GB321B UD-IQ3_XXSFits with a tweak35–70tok/sFast
Ryzen AI Max+ 395 · 128 GB321B UD-IQ3_XXSToo big—
DGX Spark · 128 GB321B UD-IQ3_XXSToo big—
RTX PRO 6000 · 96 GB321B UD-IQ3_XXSToo big—
Mac Studio with M5 Ultra · 256 GB321B UD-Q4_K_XLRuns well20–50tok/sSmooth
Mac Studio with M3 Ultra · 512 GB321B Q8_0Runs great10–20tok/sSmooth

What it is good and bad at

Good at

  • Top independent scores for coding and agents among models you can run at home
  • Only 18B parameters work per word, so it stays quick
  • MIT licence

Not good at

  • The 4-bit file is 200 GB; the 3-bit file (120 GB) is the smallest sensible one
  • Image input needs an extra 1 GB file