Meta Llama 3.2 models now available on IBM watsonx

Summary

IBM has made Meta's Llama 3.2 collection available on watsonx.ai. The collection includes the first multimodal vision models (11B/90B) and the smallest Llama variants (1B/3B). IBM positions these models for enterprise fine-tuning and hybrid cloud deployment scenarios.

Why it matters

This move brings Meta's first open multimodal models to a major enterprise AI platform. It allows organizations to access vision-language capabilities without relying on closed APIs. The 1B and 3B edge models signal a practical shift toward on-device, privacy-preserving AI. This approach could change how companies process sensitive data locally.

Highlights

• Llama 3.2 90B-Vision performs on par with GPT-4o on ChartQA. In scientific schema interpretation (AI2D) tests, it outperforms Claude 3 Opus and Gemini 1.5 Pro.
• The multimodal design adds vision parameters without altering the base large language model (LLM), keeping fine-tuning simple. The compute required for pre-training drops to approximately 2 million GPU hours, compared to 7 million GPU hours for Llama 3.1 70B.
• The 1B and 3B models target mobile and edge deployments. These models enable low-latency local inference and privacy-preserving agentic workflows.

Critical take

The benchmarking framework is selectively positive. IBM highlights gains against specific closed models but omits comparisons with the strongest current vision systems. Describing the 11B model as 'competitive' for its weight class suggests a meaningful difference in the lower segment. Furthermore, the text reads like a Meta product announcement positioning IBM as a distribution channel. The real added value of watsonx beyond model hosting is largely unaddressed.