Google's first open-weights family: Gemma 7B and 2B

On March 13, 2024, Google DeepMind published the paper behind Gemma, its first open-weights model family. Built from the research and technology used to create the Gemini models, the family ships in two sizes, 7 billion and 2 billion parameters, with both pretrained and fine-tuned checkpoints. For a company whose most capable models have been reachable only through its own APIs, that's a notable opening.

What the paper claims

The Gemma Team (Google DeepMind's research group behind the release) describes the models as "a family of lightweight, state-of-the-art open models." That state-of-the-art framing is the company's own. The concrete benchmark result in the paper: Gemma outperforms similarly sized open models on 11 out of 18 text-based tasks, across language understanding, reasoning, and safety evaluations. Two things are worth noticing about that number. The comparison class is other open models at similar sizes, not frontier closed systems, and the evaluation was run by the company that built the models.

Google is releasing both pretrained and fine-tuned checkpoints for each size. The pretrained checkpoints are meant for further training and research; the fine-tuned ones target direct use, such as instruction-following or chat. The paper says it also presents comprehensive evaluations of the safety and responsibility aspects of the models, alongside a detailed description of how they were developed.

A crowded class

March 2024 was a busy moment for the open mid-size class. Meta's Llama models had set the baseline for open instruction-tuned LLMs through 2023, Mistral had released its 7B model earlier in the year, and Microsoft had published Phi-2 in February. In that field, the interesting comparison is not "Gemma versus the biggest closed model"; it is Gemma versus the other open models you could download that same day. The paper's own framing does exactly that: 11 of 18 text-based tasks against similarly sized open models, which is the fairest comparison available in the open class.

Why it matters

The part that stands out is provenance. The Gemini models, Google's frontier systems, were trained in-house and served only through Google's infrastructure. Gemma is the first time that research has been turned into weights anyone can download and run themselves. The paper states the rationale directly: the team believes that the responsible release of LLMs is critical for improving the safety of frontier models and for enabling the next wave of LLM innovation.

For a team deciding where to run its AI workloads, the practical read is about size. A 7B model sits at the upper end of what you can serve on a single modern GPU; the 2B model is small enough to fit much more constrained hardware. If your question is whether a self-hosted model can carry a real workload (classification, extraction, summarization, a chat endpoint that doesn't need frontier reasoning), Gemma is now the reference point for what is available in the open mid-size class.

It does not replace a frontier API for the hardest tasks; that is a different class of model. And if you put Gemma in front of real traffic, run your own evaluation on your data before trusting the published numbers: they are useful, but they are the vendor's.