Gemma

Google's family of open-weight AI models you can download, run locally, and self-host

Hand-reviewed by the Toolglade team·Updated Aug 2026

chatbots#llm#open-source#google#local-ai

Free plan Claimed API Self-hosted

Toolglade’s take

One of the strongest open-weight options for developers who want Gemini-grade quality they can run privately and offline. Gemma 4's shift to an Apache 2.0 license makes it a genuinely low-friction pick for commercial and self-hosted use.

About Gemma

Gemma is Google DeepMind's family of free, open-weight AI models derived from the same research as Gemini. The latest version, Gemma 4, is multimodal (text, image, audio), supports up to a 256K context window and 140+ languages, and ships under a permissive Apache 2.0 license. Developers can download the weights, run them locally or self-host, and fine-tune them for custom applications at no cost.

Gemma is Google DeepMind's family of open-weight AI models, built from the same research and technology used to create the proprietary Gemini models. Unlike Gemini, Gemma model weights are published openly, so developers can download them, run them on their own hardware, fine-tune them on custom data, and deploy them anywhere from a laptop or phone to cloud servers. The lineup spans compact on-device sizes up to larger dense and Mixture-of-Experts variants, plus specialized members like CodeGemma, PaliGemma, ShieldGemma, MedGemma, and EmbeddingGemma. The latest generation, Gemma 4 (released March 2026), is natively multimodal with text, image, and audio input, a context window up to 256K tokens, support for 140+ languages, and function calling. It is the first Gemma release distributed under the permissive Apache 2.0 license, removing prior usage restrictions and making commercial deployment straightforward. Gemma integrates with popular tooling including Hugging Face, Ollama, LM Studio, llama.cpp, Keras, and Google AI Studio.

TL;DR

Gemma is Google DeepMind's family of free, open-weight AI models built from the same research as Gemini. The latest generation, Gemma 4 (March 2026), is natively multimodal, supports up to a 256K-token context and 140+ languages, and ships under the permissive Apache 2.0 license. Developers can download the weights and run them locally, on edge devices, or self-hosted in the cloud at no license cost. It is a leading choice for private, offline, and cost-controlled LLM deployments.

Company overview

Gemma is developed and published by Google DeepMind, Google's consolidated AI research organization, and distributed through Google's developer platform at ai.google.dev/gemma. It is the open-weight counterpart to Gemini, Google's flagship proprietary model family, and is built using the same underlying research and technology.

Since the first Gemma release in February 2024, the family has expanded rapidly into a broad ecosystem, adding generational upgrades (Gemma 2, 3, and 4) alongside specialized models such as CodeGemma, PaliGemma, ShieldGemma, MedGemma, EmbeddingGemma, RecurrentGemma, and more. This 'Gemmaverse' reflects Google's strategy of pairing its closed frontier models with a strong open-weight offering to reach developers, researchers, and on-device use cases.

Product features

Gemma provides open model weights in a range of sizes, from tiny on-device variants (such as the E2B and E4B 'effective parameter' models) to larger dense (31B) and Mixture-of-Experts (26B A4B) models, plus a 12B unified variant. Gemma 4 is natively multimodal, accepting text, image, and audio input and generating text, with a context window up to 256K tokens, multilingual support across 140+ languages, function calling, and reasoning ('thinking') capabilities.

Because the weights are open, Gemma can be downloaded, inspected, fine-tuned, and deployed anywhere. It integrates with the mainstream open-source stack, including Hugging Face Transformers, Ollama, LM Studio, llama.cpp, MLX, Keras, and Unsloth, and can be served locally, on edge hardware via Google AI Edge/LiteRT, or in production on Google Cloud and through the Gemini API. Quantization-aware trained checkpoints let larger models run on consumer GPUs.

Target market

Gemma targets developers, ML engineers, startups, researchers, and enterprises that want to run capable AI models on their own terms, whether for privacy, cost control, offline and on-device use, customization through fine-tuning, or avoiding vendor lock-in. It appeals particularly to teams building products on top of an LLM who need predictable economics and full control over deployment, as well as to the academic and open-source communities that value inspectable, tunable models.

Buyer personas

End users

Developers and ML engineers who download, prompt, fine-tune, and integrate Gemma models into applications, tools, and research workflows.

Buyers

Engineering leaders, CTOs, and technical founders who choose an open-weight model strategy for cost control, privacy, and deployment flexibility; the models themselves are free, so the decision is about engineering and infrastructure investment rather than license spend.

Key influencers

AI/platform architects, security and compliance teams, open-source and MLOps communities, and researchers who benchmark and recommend models.

Ideal customer profile

A technically capable team or organization that needs to self-host or run AI on-device for privacy, cost, latency, or customization reasons and has the engineering resources to deploy and maintain open-weight models, from indie developers and startups to enterprises with on-premise or regulated environments.

Funding & performance

part of Google

Pros & cons

Pros

Free open weights you fully own and can run offline
Gemma 4 uses a permissive Apache 2.0 license, simple for commercial use
Multiple sizes from tiny on-device models to 31B-class quality
Multimodal input (text, image, audio) and 140+ language support
Broad tooling support: Ollama, LM Studio, Hugging Face, llama.cpp, Keras
Backed by Google DeepMind and the same research as Gemini

Cons

Requires your own hardware and setup, no polished consumer app
Largest open sizes still trail top proprietary frontier models
Running bigger variants well needs a capable GPU
No managed hosting, scaling, or support out of the box
You are responsible for safety, moderation, and compliance

Pricing plans

Open weights

Free

self-host
permissive license
runs locally

Key features

API

Self-hosted

Multi-language

Integrations

Hugging Face, Google AI Studio, Ollama

Input types

text, image, audio

Output types

text, code

Best For

Developers self-hosting LLMs, On-device and edge AI, Fine-tuning on private data, Privacy-sensitive local inference

Compare key features

View all alternatives →

Feature

Gemma

Muse Glimmer

GLM (Z.ai)

Pricing

Free

Freemium

Free plan

Yes

Free trial

API

Yes

Self-hosted

Yes

Frequently asked questions

Is Gemma free?+

Yes. Gemma's model weights are published openly at no cost. You can download and run them for free, including for commercial use, paying only for whatever hardware or cloud compute you run the model on. Gemma 4 is released under the permissive Apache 2.0 license.

How does Gemma compare to Llama or Qwen?+

Gemma, Meta's Llama, and Alibaba's Qwen are all leading open-weight model families, and all can be run locally or self-hosted. Gemma is derived from the same research as Google's Gemini and is strong on multilingual and multimodal tasks; Gemma 4's Apache 2.0 license is more permissive than Llama's custom community license. The best choice depends on your size, hardware, and licensing needs, so benchmarking on your own task is worthwhile.

How do I run Gemma locally?+

The simplest way is to install Ollama or LM Studio and pull a Gemma model with a single command, giving you a local chat UI and API. Developers can also load the weights via Hugging Face Transformers, Keras, llama.cpp, or MLX. Smaller Gemma sizes run comfortably on laptops and consumer GPUs, while larger variants need more VRAM.

What license does Gemma use, and can I use it commercially?+

Gemma 4 is released under the standard Apache 2.0 license, which permits commercial use, modification, and redistribution with minimal restrictions. Earlier versions (Gemma 1 through 3) used Google's source-available Gemma Terms of Use, which also allowed commercial use but with additional prohibited-use conditions. Always review the specific model's license before deploying.

What is the difference between Gemma and Gemini?+

Gemini is Google's proprietary, closed frontier model family accessed through Google's apps and API, while Gemma is the open-weight family you can download, run locally, fine-tune, and self-host. They are built from shared research and technology, but Gemma trades some peak capability for openness, portability, and full local control.

Reviews (0)

Loading reviews…

Compare

Compare Gemma with other AI tools

Side-by-side pages for pricing, features, and best-fit use cases.

All comparisons →

Compare side by side

From the blog

All articles →

Article

AI News Roundup — Week of August 17, 2026

This week in AI: Meta's open-weight Muse Glimmer, an escalating model price war, Perplexity Comet goes free, OpenAI retires Atlas, and a wave of inference funding.

Read →

Article

The Best Local / On-Device AI Models to Run in 2026

Local AI went mainstream in 2026. A developer's guide to the best open-weight models to run on your own GPU or Mac: Muse Glimmer, Gemma, Qwen, DeepSeek, and GLM.

Read →