Gemma
Gemini's open siblings, perfect for running locally.
Olya's profile
- What sets it apart
- Gemma is Google's family of open-weight models, built on the same research behind Gemini but meant to be downloaded and self-hosted under an Apache 2.0 license. What really sets it apart is the range of sizes: from a 270M model that fits on a phone up to 27-31 billion parameters, so you choose the weight-versus-quality trade-off instead of inheriting it. Gemma 3 handles text and images, a 128K-token context and over 140 languages; Gemma 4, released in 2026, raises the context and adds audio.
- Strengths
- It runs locally on consumer hardware: a 27B Gemma 3 fits on a single 24 GB GPU like an RTX 3090, and the smaller variants reach laptops and phones. Because it is open-weight and Apache 2.0, you can use it commercially, fine-tune it on your own data and keep everything inside your walls, with no calls to an external API. It's supported everywhere that matters: Hugging Face, Ollama, llama.cpp, JAX, PyTorch and Keras.
- When to use it
- Reach for it when data must not leave your infrastructure, for privacy or compliance reasons, or when you want predictable costs instead of paying per token. It's the natural pick for offline prototyping, for fine-tuning on a specific domain, or for running something useful on edge devices where connectivity is thin.
- When to avoid it
- If you need top-tier reasoning on the hardest tasks, the closed frontier models are still ahead, and for those you may as well call the Gemini API directly. Keep in mind too that hosting a model means managing GPUs, updates and latency: if you don't want that operational load, a managed service saves you more effort than money.