Inflection Announces Inflection-2, the World's Second Most Capable Large Language Model

Inflection, has announced a significant milestone in the realm of artificial intelligence with the launch of Inflection-2. This cutting-edge model marks a remarkable stride towards realizing the vision of personal AI for everyone.

Inflection-2, unveiled today by the team at Inflection, emerges as a monumental achievement, positioning itself as the world's second most capable Language Model (LLM) within its compute class. This extraordinary accomplishment comes in the wake of Inflection's dedication to democratizing AI for widespread accessibility.

At the core of Inflection's mission lies the ambition to democratize AI, striving to make personal AI accessible to all. Merely months following the introduction of Inflection-1, the foundation of the Pi platform, Inflection-2 emerges as a leap forward, showcasing a substantial enhancement in factual knowledge, refined stylistic control, and a remarkable augmentation in reasoning capabilities.

Inflection-2’s robust performance against industry benchmarks cements its superiority. Trained on a formidable infrastructure boasting 5,000 NVIDIA H100 GPUs with ~10²⁵ FLOPs in fp8 mixed precision, this model is on par with Google’s flagship PaLM 2 Large. Inflection-2 outshines PaLM 2 Large across various standard AI benchmarks including MMLU, TriviaQA, HellaSwag & GSM8k, marking a significant stride in AI technology.

Efficiency in serving lies at the heart of Inflection-2's design, set to power Pi. Leveraging a transition from A100 to H100 GPUs and employing highly optimized inference techniques, Inflection achieves cost reduction and accelerated serving speed, despite Inflection-2's substantially larger scale.

Benchmarking against state-of-the-art models validates Inflection-2’s progress. The model’s performance across diverse benchmarks, compared to Inflection-1, LLaMA-2, Grok-1, PaLM-2, Claude-2, and GPT-4, showcases its competence across various domains.

Inflection-2 excels in MMLU, demonstrating its prowess in tasks ranging from high school to professional levels, standing out as a top performer outside of GPT-4. It showcases commendable performance in common-sense to scientific question answering benchmarks.

Surpassing expectations, Inflection-2 displays notable performance in coding tasks and mathematical reasoning, despite these not being the primary focus during its training. Further enhancements in coding capabilities are anticipated through fine-tuning on code-heavy datasets.

The arrival of Inflection-2 not only marks technological advancements but underscores Inflection's commitment to responsible AI development. As Inflection pushes boundaries, it paves the way for a future where personal AI seamlessly integrates into daily life.