from Hacker News

cerebras: 450 tokens/sec llama 3.1 70B

by davidfiala on 8/27/24, 4:10 PM with 2 comments

• by IronWolve on 8/27/24, 6:38 PM

  Cerebras fails the "how many r's in strawberry" test. Grok is the only one who passed that test.

  Going to be interesting to see the speed and accuracy keep increasing, cant imagine how fast/accurate things will be in a decade. Cant wait.

• by davidfiala on 8/27/24, 4:13 PM

  - 1,800tps on llama 3.1 8B

  - 450tps on llama 3.1 70B

  free chat interface is at: https://inference.cerebras.ai (requires login)