Google DeepMind
·
8 months ago
Google introduced T5Gemma, a collection of encoder-decoder language models created by converting pretrained decoder-only Gemma 2 models into encoder-decoder architectures through a technique called adaptation. T5Gemma models range from 2B to XL sizes, with a 9B-2B variant that pairs a large encoder with a small decoder to optimize quality-efficiency trade-offs. The adapted models outperform their decoder-only equivalents on reasoning tasks—for example, T5Gemma 2B-2B instruction-tuned achieves a 12-point improvement on MMLU over Gemma 2 2B—and the pretrained and fine-tuned checkpoints are now available on Hugging Face and Kaggle.
Google DeepMind
·
8 months ago
Google released MedGemma 27B Multimodal and MedSigLIP, two new medical AI models designed for healthcare applications involving text, images, and electronic health records. MedGemma 27B scores 87.7% on the MedQA medical knowledge benchmark while requiring approximately one-tenth the inference cost of DeepSeek R1, and can run on a single GPU. Developers can now fine-tune these open models for specific medical tasks like report generation and image classification while retaining full control over privacy and infrastructure.
Google DeepMind
·
8 months ago
Google released Gemma 3n, a mobile-focused AI model family designed to run multimodal applications directly on edge devices with minimal memory requirements. The E4B version achieves an LMArena score over 1300, making it the first model under 10 billion parameters to reach this benchmark, while requiring as little as 2GB to 3GB of memory for operation. The model's MatFormer architecture enables developers to extract custom-sized variants between two pre-built sizes, with the E2B sub-model offering up to 2x faster inference than the larger E4B version.
Google DeepMind
·
8 months ago
Google released the stable version of Gemini 2.5 Flash-Lite, a lightweight model designed for fast, cost-efficient inference across tasks like translation and classification. The model costs $0.10 per million input tokens and $0.40 per million output tokens, with a 1 million-token context window and support for native tools including grounding with Google Search and code execution. Users can deploy the model by specifying "gemini-2.5-flash-lite" in their code, with Google retiring the preview alias on August 25th.
