Llama-3.1 Nemotron Ultra 253B vs UI-TARS-72B-DPO

Llama-3.1 Nemotron Ultra 253B

NVIDIA

Llama-3.1-Nemotron-Ultra-253B-v1 is a 253-billion-parameter model from NVIDIA, derived from Meta's Llama 3.1 405B using neural architecture search (NAS) compression combined with NVIDIA's Nemotron post-training pipeline, which recovers and exceeds the base model's capability after structural compression. Released in April 2025, it supports toggling between a standard instruction mode and an extended reasoning mode via system prompt, allowing the same model to handle both rapid responses and deliberate chain-of-thought tasks. It is the flagship of the Nemotron family, available open-weight on HuggingFace and through NVIDIA NIM for enterprise inference.

ByteDance

UI-TARS-72B-DPO, released by ByteDance in early 2025, is a 72 billion parameter multimodal large language model from the UI-TARS family, built on Qwen-2-VL and fine-tuned for automated GUI interaction and computer control. It features native understanding of screenshots, UI elements, and web interfaces, achieving strong results across GUI benchmarks for perception, grounding, and agentic control. UI-TARS-72B-DPO targets computer-use agents, web automation, and applications requiring robust visual UI reasoning.