Cerebras beats Nvidia Blackwell on Llama 4 Maverick: 2,522 tokens/s
In one sentence On May 28, 2025 Cerebras announces that Artificial Analysis measured its Llama 4 Maverick (400B) endpoint at 2,522 tokens/s per user, versus 1,038 for Nvidia Blackwell, a record Nvidia had announced only a week earlier.
Cerebras makes "wafer-scale" chips: instead of cutting a silicon wafer into many processors, it uses a whole wafer as one huge chip, with lots of fast memory directly on board. Its strength is inference, making models respond very quickly.
One week after Nvidia claimed about 1,000 tokens per second per user on Llama 4 Maverick with 8 Blackwell GPUs, Cerebras announces it has done better: the independent benchmarking firm Artificial Analysis measured its service at 2,522 tokens per second, against 1,038 for Blackwell.
In the same comparison the other providers measured were much slower: SambaNova 794, Groq 549, Amazon 290, Google 125, Microsoft Azure 54 tokens/s. The figures come from Cerebras' press release, which quotes the CEO of Artificial Analysis.
Companies
Cerebras Systems, Artificial Analysis
Tools
Cerebras Inference, WSE-3, Llama 4 Maverick
Tags
Sources