Multimodal

TL;DR for operators A DPO-style preference pair usually contains one prompt and two ranked responses. Conventional execution turns that into two prompt-response sequences, which means the same prompt is processed twice. Jaekyung Cho’s Preference Packing: Efficient Preference Optimization for Large Language Models1 treats that duplication as a systems problem. It stores the prompt once, places the alternative responses behind it, and uses masking plus adjusted position IDs so each response still behaves as though it were paired independently with the prompt. ...

An open-weight vision encoder developed by Cohere For AI, part of Project Aya’s global multilingual and multimodal research initiative.

OpenAI’s flagship GPT-4 variant that natively supports text, vision, and audio input/output with faster performance and improved reasoning.

A mid-sized member of Anthropic’s Claude 3 model family, optimized for balanced performance across reasoning, speed, and multimodal understanding.

A widely used multimodal model from OpenAI that learns joint image–text embeddings, enabling zero-shot image classification, search, and multimodal applications.

A multi-modal foundation model by DeepSeek AI, integrating vision and language for high-performance tasks including OCR, captioning, and visual reasoning.

A fast and lightweight variant of Google’s Gemini 2.0 multimodal model, optimized for latency-sensitive tasks via the Gemini API.

The next-generation model from xAI, built on a new architecture and fully integrated into X (formerly Twitter) as part of Elon Musk’s AI assistant efforts.

A next-generation vision-language model by Google, combining Gemma LLM and SigLIP vision encoder for image captioning, VQA, and image-text reasoning tasks.