The AI Bench
Liquid AI · LFM Open 1.0

LFM2.5

Liquid AI’s small on-device models, tuned for tool use rather than coding or knowledge.

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
1.2BMXFP40.7 GB1.8 GB2.1 GB—
2.7BQ4_K_M1.7 GB2.8 GB3.2 GB4.8 GB
Q8_02.9 GB4.0 GB4.4 GB6.0 GB
8.5B (1.5B active)Q4_K_M5.2 GB6.3 GB6.6 GB—
Q8_09.0 GB10.1 GB10.4 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 GB8.5B Q8_0Runs great19–40tok/sSmooth
RTX 4060 · 8 GB8.5B Q4_K_MRuns well75–150tok/sFast
RTX 3060 · 12 GB8.5B Q8_0Runs well65–120tok/sFast
RTX 5060 Ti · 16 GB8.5B Q8_0Runs great75–140tok/sFast
RTX 4090 · 24 GB8.5B Q8_0Runs great120–240tok/sFast
RTX 5090 · 32 GB8.5B Q8_0Runs great160–310tok/sFast
Mac with M4 · 16 GB8.5B Q8_0Runs well30–60tok/sFast
Mac with M4 Pro · 48 GB8.5B Q8_0Runs great45–100tok/sFast
Mac with M5 Max · 128 GB8.5B Q8_0Runs great65–140tok/sFast
Ryzen AI Max+ 395 · 128 GB8.5B Q8_0Runs great50–95tok/sFast
DGX Spark · 128 GB8.5B Q8_0Runs great50–100tok/sFast
RTX PRO 6000 · 96 GB8.5B Q8_0Runs great160–310tok/sFast
Mac Studio with M5 Ultra · 256 GB8.5B Q8_0Runs great65–140tok/sFast
Mac Studio with M3 Ultra · 512 GB8.5B Q8_0Runs great55–120tok/sFast

What it is good and bad at

Good at

  • Fast on small devices
  • Tool use

Not good at

  • Not for coding or knowledge-heavy work (their own card says so)
  • Free for business only under $10M yearly revenue

Start it

Install Ollama, then run one of these in a terminal. LM Studio users can search the same name in the app.

ollama run lfm2.5:8b