AI chip startup Cerebras outperforms NVIDIA’s Blackwell in Llama 4 test

Cerebras has achieved a groundbreaking milestone in AI inference performance, establishing a new world record for processing speed with Meta’s flagship large language model. By delivering over 2,500 tokens per second on the massive 400B parameter Llama 4 Maverick model, Cerebras has demonstrated that specialized AI hardware can significantly outpace even the most advanced GPU solutions, reshaping performance expectations for enterprise AI deployments.

The big picture: Cerebras has set a world record for LLM inference speed, achieving over 2,500 tokens per second with Meta’s 400B parameter Llama 4 Maverick model.

• Independent benchmark firm Artificial Analysis measured Cerebras at 2,522 tokens per second per user, more than doubling the 1,038 tokens per second recently announced by NVIDIA’s flagship Blackwell GPUs.
• This performance milestone makes Cerebras the first and only inference solution to break the 2,500 TPS barrier with the largest and most powerful model in the Llama 4 family.

Why this matters: Inference speed is critical for advanced AI applications, particularly for agents, code generation, and complex reasoning tasks where responsiveness determines usability.

By the numbers: Artificial Analysis tested multiple vendors against the same model, revealing a significant performance gap:

• Cerebras: 2,522 tokens per second
• NVIDIA Blackwell: 1,038 tokens per second
• SambaNova: 794 tokens per second
• Groq: 549 tokens per second
• Amazon: 290 tokens per second
• Google: 125 tokens per second
• Microsoft Azure: 54 tokens per second

What they’re saying: “Cerebras is the only inference solution that outperforms Blackwell for Meta’s flagship model,” said Micah Hill-Smith, Co-Founder and CEO of Artificial Analysis.

Competitive advantage: Cerebras claims its solution is optimized specifically for Llama 4 and is available immediately without requiring special software optimizations.