Category	Benchmark	Gemini Ultra	GPT-4 / GPT-4V	Vision (Rose)
General	MMLU	90.0%	86.4%	86.0%
Reasoning	Big-Bench Hard	83.6%	83.1%	84.1%
	DROP	82.4	80.9%	85.2%
	HellaSwag	87.8%	95.3%	89.6%
Math	GSM8K	94.4%	92.0%	89.1%
	MATH	53.2%	52.9%	58.9%
Code	HumanEval	74.4%	67.0%	84.7%
	Natural2Code	74.9%	73.9%	86.5%
Image (Multimodal)	MMMU	59.4%	56.8%	63.1%
	VQAv2	77.8%	77.2%	82.2%
	TextVQA	82.3%	78.0%	91.8%
	DocVQA	90.9%	88.4%	93.2%
	Infographic VQA	80.3%	75.1%	90.9%
	MathVista	53.0%	49.9%	67.6%

Reasoning Performance

Multimodal Reasoning in Speech

ROSE's reasoning advantage comes from its multimodal foundation:

• Contextual Understanding: Processes not just words, but meaning, intent, and emotional context.

• Persistent Memory: Maintains reasoning threads across long conversations

• Cultural Intelligence: Reasons about cultural nuances and appropriate responses

• Real-time Inference: Applies complex logical reasoning while generating speech

““we are outpacing OpenAI, Google, and Meta with a better model, built on far less compute and at a fraction of the cost.””

The Breakthrough

ROSE represents the first speech-to-speech model backed by foundational reasoning intelligence. By inheriting Vision AI's proven reasoning capabilities, outperforming both GPT-4 and Gemini across multiple cognitive benchmarks, ROSE doesn't just sound human, it thinks and reasons like one. This foundation enables conversations that are not only natural but genuinely intelligent.