About Plain language, no marketing

About Gemma 4

What it is, who made it, why people care, and what this reference is for — written for someone arriving with no prior context. If you already know what an open-weight model is, the datasheet will serve you better.

The short version

What Gemma 4 is

📦

A file you can download

Not a service you log into

Gemma 4 is a set of model weights — large files of learned numbers that a program on your computer reads to generate text. Google publishes them openly, so you download them once and run them yourself. There's no account, no per-message cost, and no requirement to be online.

This is the fundamental difference from ChatGPT, Claude or Gemini, which run on someone else's servers and are reached over the internet.

🏢

Made by Google DeepMind

Released March 2026

Built from the same research as Gemini 3, Google's closed flagship, then released under Apache 2.0 — a standard open-source licence that permits commercial use, modification and redistribution without conditions.

Gemma is the open sibling of Gemini rather than a cut-down version of it. The two share lineage but are distributed on opposite terms.

💡
The one-sentence version. Gemma 4 is Google's family of free, downloadable AI models that run on your own hardware — from a phone up to a server — and that you can use commercially without asking anyone's permission.
Why it matters

Four things that follow from being open

Not abstract benefits — each one changes what you can build.

🔒

Your data stays put

Prompts never leave the machine. For confidential documents, medical notes or anything under a data-residency rule, that's a guarantee you can verify by unplugging the network.

💸

No per-message cost

Once downloaded, running it costs only electricity. There's no meter, and a traffic spike doesn't become an invoice.

📴

Works offline

On a plane, in a lab with no outbound network, or on a phone in airplane mode. The model doesn't know or care whether you're connected.

🔧

You can change it

Fine-tune it on your own data, quantise it, embed it in a product, and redistribute the result. Apache 2.0 imposes no conditions on any of that.

⚖️
And the honest counterweight. Open weights also mean no provider filtering harmful outputs, no one to call when something breaks, and a model that's behind the frontier on the hardest reasoning. Gemma 4 isn't the most capable model available — it's the most capable one you can own. Whether that trade is right depends entirely on what you're building.
The family

Five models, one design

They share a tokenizer, a chat format and a tool-calling interface, so moving between them is a configuration change rather than a rewrite.

ModelRuns onBest for
E2B 2.3BA phone Offline assistants, classification, extraction — 1.1 GB as a mobile build
E4B 4.5BA laptop The same, with more headroom; 4.5 GB at 4-bit
12BOne consumer GPU The all-rounder — and the only large model that accepts audio and video
26B A4B mixture of expertsA 24 GB card Serving many users — roughly 4.4× faster than the 31B
31BA 24 GB card Maximum quality when that decides the outcome
🎧
One counterintuitive detail worth knowing early. The biggest model is not the most capable in every respect. Audio and video input live in the smaller models — the 12B and the two edge models — because that's where the encoder-free architecture landed. If your task involves sound, the 12B is your ceiling, not the 31B.
How it got here

Four generations in two years

February 2024
Gemma 1
2B and 7B, text only, 8K context. Google's first open-weight release in the Gemini era.
June 2024
Gemma 2
2B, 9B and 27B. Still text, still 8K.
March 2025
Gemma 3
Vision arrives, along with a 128K context and function calling. Sizes from 270M to 27B, later joined by a deliberately tiny 270M model for task-specific fine-tuning.
June 2025
Gemma 3n
A separate on-device family, E2B and E4B, adding audio.
March 2026
Gemma 4 Current
The largest jump so far. Reasoning improves dramatically — competition maths goes from 20.8 to 89.2 — context doubles to 256K, audio and video arrive properly, and the licence changes to Apache 2.0. The 12B follows in June, and a weights refresh in July fixes tool calling.
📈
The generational jump was training, not size. Gemma 4's models are similar in scale to Gemma 3's, yet a 2-billion-parameter Gemma 4 now outscores the previous generation's 27B flagship on reasoning and code. Almost all of that came from reasoning-focused post-training and a "thinking mode" that lets the model work through a problem before answering.
Where it sits

Gemma 4 among the alternatives

Positioning, stated plainly rather than favourably.

🏔️

Against frontier closed models

Gemini 3, GPT-5 and Claude are ahead on the hardest reasoning, and that gap is real. Gemma's answer isn't that it competes — it's that you can hold it.

🤝

Against other open models

Competitive, and unusually strong at the small end. Alibaba's Qwen line is the closest rival and its 2.4-trillion-parameter flagship outscores Gemma 4 — but needs a cluster to run.

📱

At the bottom end

Where Gemma 4 is genuinely unmatched. A 2B model scoring 37.5 on competition maths, in under a gigabyte on a phone, has no real equivalent in the open ecosystem.

About this reference

What this site is, and isn't

What it is

An independent reference

  • ✅ Specifications, benchmarks and provenance in one place
  • ✅ Practical guides for downloading, installing and self-hosting
  • ✅ Interactive tools — a VRAM calculator, runtime selector and model comparator
  • ✅ Honest coverage of limitations, not only strengths
  • ✅ Every figure traceable to a named source
What it isn't

Not affiliated with Google

  • ⚠️ Not operated, endorsed or reviewed by Google or Google DeepMind
  • ⚠️ Not a substitute for the official documentation
  • ⚠️ Not legal advice on licensing, however carefully it's summarised
  • ⚠️ Not guaranteed current — this field moves weekly

Gemma and Google are trademarks of Google LLC. Always confirm anything consequential against ai.google.dev/gemma.

How the figures were sourced

Type of claimWhere it comes fromHow it's treated
Architecture and benchmarksGemma 4 technical report, model cards Marked as self-reported; evaluation settings stated
Model IDs and file sizesGoogle's documentation and repositories Checked against the live repositories, not transcribed
Throughput and agentic resultsThird-party benchmarking Marked as independent — and often less flattering
Derived figuresComputed here Formula shown, and validated against published measurements where they exist
Anything unavailable— Left blank and said so, rather than estimated silently
🔍
Where sources disagreed, the page says so. Google's published memory figures run above raw parameter arithmetic; LMArena rankings quoted at launch differ from those in the technical report because ranks decay; and the widely repeated claim that MLX is faster than GGUF on Apple silicon doesn't hold for Gemma 4. In each case both figures appear with the reason for the difference, rather than one being quietly chosen.
Where to go next

Depending on why you're here

Just want to try it→
Runtime selector

Four questions and you'll have the right runtime, model size and exact commands for your machine.
Will it fit?→
VRAM calculator

Weights, KV cache, context and concurrency, with a straight answer on which GPUs have room.
Need the numbers→
Full datasheet

Every published figure for the family, with a provenance table for each one.
Have a question→
FAQ

134 questions, searchable, each linking to the page with the full working.
Evaluating it→
Limitations

What it does badly, with a mitigation for each — worth reading before you commit.
Ready to go→
Downloads

Every checkpoint with its exact repository ID and pull command.

An open model you can actually own.

Free weights, five sizes, and hardware requirements that start at a phone.

Download Gemma 4 Official docs Technical report