SambaNova Launches Fastest AI Platform with Record 132 Tokens/Sec for Llama 3.1 405B

Sep 10, 2024, 02:48 PM

SambaNova has launched its new cloud inference platform, SambaNova Cloud, which is now available for developers to access Llama 3.1 models including 8B, 70B, and 405B on their custom AI chips. The platform sets a new record for inference speed, achieving 132 tokens per second for Llama 3.1 405B at full precision and 570 tokens per second for Llama 3.1 70B. This performance is 10 times faster than traditional GPUs. The API is available for free with no waitlist, enabling developers to unlock advanced AI applications. Additionally, Llama 3.1 405B can also achieve 100 tokens per second on TogetherCompute API, with a 128k long-context version coming soon.