Google also released a tabular foundation model, rivaling TabPFN. However, it comes with a non-commercial license.
A conversion of the 26B-A4B Gemma model into a diffusion model, which makes it attractive for low batch sizes, i.e., local usage at home, for example, on a MacBook or a Spark.
A music generation model by Google. This kind of model is really rare in the open model space.
An encoder-free VLM of the Gemma series, which (partially due to its license change to Apache 2.0) has seen rapid adoption.
The long-awaited update to the Gemma series, featuring multiple sizes: 4B, 9B, and 31B dense models, as well as a 26B-A4B MoE. Even more importantly, with Gemma 4, Google has decided to use Apache 2.0 as its license, which removes the uncertainty and legal challenges around interpreting custom licenses.
The long-awaited update to the Gemma series, featuring multiple sizes: 4B, 9B, and 31B dense models, as well as a 26B-A4B MoE. Even more importantly, with Gemma 4, Google has decided to use Apache 2.0 as its license, which removes the uncertainty and legal challenges around interpreting custom licenses.
The long-awaited update to the Gemma series, featuring multiple sizes: 4B, 9B, and 31B dense models, as well as a 26B-A4B MoE. Even more importantly, with Gemma 4, Google has decided to use Apache 2.0 as its license, which removes the uncertainty and legal challenges around interpreting custom licenses.
The long-awaited update to the Gemma series, featuring multiple sizes: 4B, 9B, and 31B dense models, as well as a 26B-A4B MoE. Even more importantly, with Gemma 4, Google has decided to use Apache 2.0 as its license, which removes the uncertainty and legal challenges around interpreting custom licenses.
A fine-tuned version of Gemma for, well, translations. The base Gemma is sort of an insider tip in terms of its strong multilingual abilities, so expect this version to be even better for those tasks.
A small model for function calling.
A Gemma version trained with differential privacy, which is important for certain sectors and applications, such as healthcare.
A tiny version of Gemma for debugging and maybe some automation tasks (models this small historically haven't been coherent enough to do much).
A tiny embedding model by Google supporting Matroska representation, i.e., different dimensional embeddings.
Yes, it is 2025 and we are getting a modern version of T5 (Text-To-Text Transfer Transformer, 2019, encoder-decoder models). Some of the versions use Gemma 2 (2B and 9B) for the decoder part, while others are using the same sizes as mT5 and are trained from scratch, allowing an easy switch between old deployments still using T5 and those new versions. Considering that T5 still rakes in millions of downloads each month, those models could end up being similarly popular. However, these new versions are licensed under Gemma, not under Apache 2.0 like the original T5 models.
An instrumental-only music-generation model.
An encoder for videos.
A Gemma variant for medical texts.
A version of Gemma which uses Per-Layer Embeddings to speed up the inference process. This allows the model to use less than half the memory of the original model, making it suitable for edge devices and smartphones.
A fine-tuned version of Gemma for therapeutic development.
Aside from the points mentioned in our post, the Gemma release highlights some tricks used by labs: Knowledge Distillation using a big teacher model, a (5:1) local / global attention layer ratio. The latter is a configuration outlined by Noam Shazeer during his time at Character.AI. Apart from the instruction models, Google also releases the pre-trained base model.
Google has dropped quantization-aware trained versions of Gemma, a technique to bring the performance of the int4 quantized models close to the original model. The 27B version of Gemma only needs 14GB of VRAM instead of the 54GB of the original model.
A Gemma-based model for classifying images based on various safety categories. Although these models don't get a lot of attention, they are useful when deploying (AI) services in production.
An update to the popular SigLIP models, improving the performance across the board.
Updated versions of the PaliGemma 2 model series on more tasks like OCR.
A Japanese-focused version of Google's Gemma models.
A funny RAG version of Gemma from Google.
This is a serious model. I could write a speculative post about each of the sections in the report. In summary, it evaluated on ChatBotArena well, is trained on LMSYS data, is distilled similarly to Gemini (probably, as discussed in my recent post), uses model merging during fine-tuning, uses an order of magnitude larger reward model for RLHF (>100B parameters), uses synthetic and human data, and is a reasonable size for inference on one 80GB memory GPU. Read more in the technical report here. Otherwise, I seriously expect future Gemma models to replace a lot of Llama models in workflows. Google shows every intention of putting a lot of weight behind these, which is fantastic to see. Hopefully it can continue. For more on Gemma 2, see this post from HuggingFace.
Google release a very solid visual language model in its Gemma suite. Folks online have been impressed - this space is really heating up!
A cool transformer focusing on time-series data (such as weather forecasting?). Seems good for science.