UL2 vs ggml.ai

UL2 is a unified framework for pre-training language models that perform well across a wide range of natural language processing tasks. It separates model architecture from training objectives, allowing flexible combinations of self-supervised learning methods. The core innovation is the Mixture-of-Denoisers (MoD) objective, which blends multiple denoising tasks to improve generalization. UL2 introduces mode switching, linking downstream fine-tuning to specific pre-training modes for better task adaptation. Scaled up to 20 billion parameters, UL2 achieves state-of-the-art results on over 50 NLP benchmarks, including language understanding, generation, reasoning, and knowledge grounding. It also excels at in-context learning, outperforming larger models like GPT-3 on zero-shot and one-shot tasks. The framework supports instruction tuning (Flan-UL2), further enhancing performance on complex reasoning and multitask benchmarks. Open-source Flax-based T5X checkpoints for UL2 and Flan-UL2 20B models are publicly available, facilitating research and application development.

ggml runs large language and speech models on everyday CPUs and GPUs through a compact C tensor library built for on-device inference. ML engineers and app developers adopt it via llama.cpp and whisper.cpp when they want LLaMA or Whisper workloads without cloud-only dependencies.

Frameworks like PyTorch optimize for training clusters and heavy runtimes. ggml keeps the core library minimal with zero runtime memory allocations, no third-party dependencies, and integer quantization so llama.cpp can serve Meta LLaMA weights on laptops and Apple Silicon.

The ggml.ai company was founded in 2023 by Georgi Gerganov to support the library and was acquired by Hugging Face in 2026. The core ggml project stays MIT licensed with open development on GitHub.