Meta launches the Llama API with record inference speeds for developers

In an exciting turn of events in the world of artificial intelligence, Meta has unveiled the Llama API at the first LlamaCon, promising to revolutionize the way developers interact with their AI models. This new service, which is in a phase of limited free trial, allows developers to access different models from the Llama family, including the newly released Llama 4 Scout and Llama 4 Maverick.

The Llama API stands out for its ease of use, offering API key creation with a single click and lightweight SDKs in TypeScript and Python. Best of all, it is compatible with the OpenAI SDK, making it easier for developers to port their OpenAI-based applications to this new platform.

Unprecedented inference speeds

But that’s not all, as Meta has joined forces with Cerebras and Groq, promising record inference speeds. Cerebras claims that its Llama 4 Cerebras model can generate tokens up to 18 times faster than traditional NVIDIA GPU-based solutions and others. According to the benchmark site Artificial Analysis, the Cerebras model exceeded 2,600 tokens/s for Llama 4 Scout, compared to only 130 tokens/s from ChatGPT and 25 tokens/s from DeepSeek.

Andrew Feldman, CEO and co-founder of Cerebras, expressed his enthusiasm: “Cerebras is proud to make the Llama API the fastest inference API in the world. Developers building real-time applications need speed. With Cerebras in the Llama API, they can create AI systems that are fundamentally unattainable for leading GPU-based inference clouds.”

Interested developers can access this incredible inference speed by selecting Cerebras from the model options within the Llama API. Additionally, Llama 4 Scout is also available through Groq, although it currently operates at over 460 tokens/s, which is approximately 6 times slower than the Cerebras solution, but still 4 times faster than other GPU-based solutions.