CBRSPrivate
Evidence
96%Authoritative
FactConfirmedProduct·June 2, 2025

Cerebras Sets Llama 4 Maverick Inference Speed Record at 2,500+ Tokens/Sec

InferenceWSE Chip

Cerebras achieved a world record of 2,500+ tokens/sec per user on Meta's 400B parameter Llama 4 Maverick model as measured by Artificial Analysis, more than doubling Nvidia's DGX B200 benchmark of 1,000 tokens/sec.

Evidence Strength

Evidence
96%Authoritative
Backed by official company doc
Single publisher source
Includes official or primary source
Key Development
High-significance development (rated 8/10)
Confirmed — verified event

Insights

First tracked

June 2, 2025

Last updated

June 2, 2025

Sources

1 source

Related Developments

CS-3 vs. NVIDIA DGX B200 Blackwell Benchmarks PublishedCerebras Delivers 3,000 Tokens/Second Inference for OpenAI's gpt-oss-120B Open-Weight ModelJais 2 Arabic-Centric LLMs Trained and Deployed on Cerebras Wafer-Scale ClustersGLM-4.7 Available on Cerebras Inference Cloud at 1,000-1,700 Tokens/SecondOpenAI Signs $10B+ Multiyear Compute Deal with Cerebras

Sources (1)

Source Timeline

View all Cerebras Systems developments
Cerebras Sets Llama 4 Maverick Inference Speed Record at 2,500+ Tokens/Sec — Cerebras Systems | OpenCall