174 points57 comments
Cerebras Runs Trillion-Parameter LLMs at 2K Tokens/sec
What is Cerebras Deploys Trillion-Parameter AI Models?
Cerebras splits a 3-trillion-parameter model across multiple wafer-scale chips, using 4-bit quantization to fit the weights into on-chip memory. The system delivers inference speeds of 1,000, 2,000+ tokens per second.
Who it is for
Anyone needing frontier-scale language models for research or product prototypes can now get near-real-time responses without a massive GPU farm.
How to try it
Visit cerebras.ai/wafer-scale or contact sales for access.