We’re excited to announce three major updates to Meta Llama models on Oracle Cloud Infrastructure (OCI) Generative AI service.
Updates for Llama Models on OCI
First, Meta’s Llama 3.3 70B model is now generally available. This text-only model delivers better performance than both Llama 3.1 70B and Llama 3.2 90B for text tasks. It uses the same prompt format as 3.1 70B and maintains the 128k context length and code interpreter support.
Second, we’ve released a new FP8 quantized version of Meta’s frontier model, Llama 3.1 405B, with a 50% reduced GPU footprint. You can now deploy new dedicated compute instances more cost effectively while maintaining performance.
Third, our OCI GenAI team has added tool support to Llama models, enabling them to integrate with external tools and APIs to enhance their capabilities. This integration allows the models to fetch real-time data, run code, and interact with databases, making their responses more accurate and contextually relevant. This feature broadens the models’ utility, enabling them to handle complex queries and perform tasks requiring external data or processing.
Highlights of these updates include the following examples:
- Llama 3.3 70B is a text-only model
- Llama 3.3 70B uses the same prompt format as Llama 3.1 70B
- Llama 3.3 70B supports the same code-interpreter and retains the 128k context length
- Llama 3.3 70B improvements include reasoning, coding, math, and instruction-following. For details, see the Llama 3.3 model card.
- Llama 3.3 70B is available for on-demand and dedicated hosting
- Tools support for Llama models
- New dedicated instances of Llama 3.1 405B FP8 require half the GPU compute
Oracle Cloud Infrastructure customers can use models without infrastructure management concerns. Access is available through chat interfaces, APIs, or dedicated endpoints.
Want to Know More?
For integration details and pricing information, consult the Generative AI service documentation or contact your Oracle representative.
For more information see the following resources: