Gemma 4 is Google DeepMind's open-weight model family, from tiny edge-friendly variants (E2B, E4B) to a 31B dense model, free to download and licensed for commercial use, with a 128K-256K context window depending on size. Pick it when you want to run an open model locally or self-hosted; pick a hosted model when you'd rather not manage the hardware.
Other good fits
A project that needs commercially licensed open weights, not just a hosted API you don't control.
+ Pros
Read pros
Gemma 4 is a family of open-weight models from Google DeepMind, spanning ultra-efficient edge models (E2B, E4B) through a unified 12B multimodal model, a 26B mixture-of-experts model (4B active parameters per token) and a 31B dense model for server-grade use. The weights are free to download and licensed to permit responsible commercial use, so you can tune and deploy them in your own projects without a per-token API fee.
Pick it when…
You want to run a capable model locally or on your own server, choosing a size that actually fits your hardware.
You need commercially licensed open weights, not just access to a hosted API.
You want to fine-tune or adapt the model for a specific task rather than relying only on prompting a hosted model.
− Cons
Read cons
The smallest, most hardware-friendly variants trade capability for that efficiency; don't expect edge-model performance to match a much larger model on hard reasoning tasks. Running any variant locally means you own the setup, updates and hardware costs that a hosted API would otherwise absorb into its per-token price.
Pick something else when…
You don't want to manage hardware or a runtime at all. A hosted model through the Claude API or a similar provider removes that entirely.
You need frontier-level reasoning on the hardest tasks. Larger hosted models still lead there; Gemma 4 is strongest for efficient, self-hosted use.
You want a runtime that makes trying different open models easy.LM Studio or Ollama are built for exactly that, including Gemma 4.
Avoid if
You need the absolute largest context window or frontier-level reasoning; larger hosted models still lead on the hardest tasks.
Context
Smaller Gemma 4 models (E2B, E4B) offer a 128K-token context window; the medium models (12B and up) support 256K. The 26B mixture-of-experts model activates only 4B parameters per token during inference, but still needs to load all 26B parameters into memory, so its memory footprint is closer to a dense 26B model than its "active" parameter count suggests.
Requirements
Running Gemma 4 locally needs a runtime (Ollama, LM Studio or similar) and hardware matching the chosen model size, from a modest laptop for the smallest edge variants up to a serious GPU for the 31B model. No account or API key is needed for local use; a hosted provider serving Gemma 4 would have its own access requirements.
Licence & cost
The model weights themselves are free; the real cost is hardware. In BF16 precision, the E2B model needs roughly 11.4GB of GPU/TPU memory (or about 1.1GB on mobile with appropriate optimisation), while the 31B model needs around 69.9GB. Quantised versions (4-8 bit) cut those requirements substantially, making even the larger models runnable on more modest hardware at some quality cost. Running Gemma 4 through a hosted provider (rather than locally) shifts the cost to that provider's per-token rate instead of your own hardware.
First task
Documentation-based suggestion — not run here.
Download the smallest variant that fits your hardware (E2B or E4B) and run it locally through a runtime like Ollama or LM Studio on a made-up summarisation task, then compare its output against a larger hosted model on the same prompt.
Done when you can say, from an actual comparison, whether the smaller local model is good enough for your task or whether you need a larger Gemma variant or a hosted model instead.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Rests on Google's Gemma model documentation, read on 2026-09-23. No Gemma 4 model was downloaded or run for this review.
Run open models locally with text and image inputs.
Choose a model size that fits a phone, laptop, or workstation.
Use GGUF or MLX formats for supported local runtimes.