Inflection AI Introduces Inflection-2, Outperforms Llama 2 and PaLM 2

Inflection AI, the startup which created  conversational chatbot Pi, has introduced a new AI model named Infection-2. The company asserts that this model has the capability to outperform two well-known alternatives developed by Google and Meta, positioning itself closely behind OpenAI's larger flagship model, GPT-4.

Inflection-2 is trained on 5,000 NVIDIA H100 GPUs in fp8 mixed precision, amounting to approximately 10²⁵ FLOPs. This newly-released model is set to be integrated into Pi, the chatbot introduced by Inflection in May. Inflection maintains close partnerships with Microsoft, Nvidia, and CoreWeave for the management of its compute cluster. 

Inflection-2 outperformed the largest, 70 billion parameter version of LLaMA 2, Elon Musk’s xAI startup Grok-1, Google’s PaLM 2 Large, and startup Anthropic’s Claude 2, trailing only behind GPT-4 in the MMLU task.

The company stated that, despite math and code benchmarks not being explicit focuses, Inflection-2 performed well across four of them. However, on the two benchmarks for which OpenAI has shared results, it notably trailed GPT-4 by a considerable margin.

It also performed best on two of three questions-and-answers task benchmarks, losing to PaLM 2 Large in one. 

Inflection-2 offers improved cost-effectiveness and speed compared to Inflection-1, despite its larger size. The transition from A100 to H100 GPUs, coupled with a highly optimized inference implementation, contributed to this achievement. Moreover, Inflection AI looks forward to training even larger models using the full capacity of our 22,000 GPU cluster.