Gemma 2
#32google · v2 · seit 27. Juni 2024 (9B und 27B); Gemma 2 2B später im Jahr 2024 · 10× · zuletzt 30. Juni 2026
16
Momentum
Gemma 2 is a family of open, lightweight language models from Google DeepMind, built on the same research and technology as the Gemini models. It was released on June 27, 2024 in 9B and 27B sizes, with a 2B variant added later. The models are text-to-text, decoder-only models with open weights (base and instruction-tuned variants), optimized for inference speed and efficiency across diverse hardware, from laptops to cloud TPUs/GPUs. Gemma 2 is distributed under the commercially-friendly Gemma license and available via Kaggle, Hugging Face, and Google Cloud Vertex AI.
Historique du momentum
04.04.03.07.
Fonctionnalités
| Key Benchmark (%) | MMLU 27B (instruction-tuned): 76.2%; 9B (IT): 72.3%; 2B (IT): 56.1% |
| Context Window (Tokens) | 8K tokens (8192), with sliding window attention (local 4096 tokens in every other layer) |
| License | Gemma License (commercially usable, not a full open-source standard) |
| Multimodality | No – text-to-text models only (input: text, output: text); image capability only via derived model PaliGemma 2 |
| Platform | Hugging Face, Kaggle, Google AI Studio, Vertex AI, Ollama, Hugging Face Transformers, JAX, PyTorch, TensorFlow/Keras, vLLM, Gemma.cpp, Llama.cpp, NVIDIA TensorRT-LLM/NIM |
| Release Date | June 27, 2024 (Gemma 2 9B/27B); 2B variant followed later in 2024 |
