Gemma 4

Google's open-weight Gemma 4 family — Apache 2.0 models from a 2.3B-effective edge model to a 30.7B dense flagship, with QAT builds for local deployment.

Last reviewed September 6, 2026

At a glance

Overview

Gemma 4 is Google DeepMind’s current open-weight model family. Google’s release notes date the first checkpoints to March 31, 2026; the launch blog post — and the model’s appearance in the Android AICore Developer Preview — followed on April 2, 2026. The family has been extended twice since. It is licensed under Apache 2.0 verbatim, not a bespoke community licence, so commercial use, modification and redistribution are unrestricted.

Google’s framing is intelligence density: the family is “purpose-built for advanced reasoning and agentic workflows,” and the headline claim from the launch post is that Gemma 4 outperforms “models 20x its size.” The evidence offered is leaderboard placement — the 31B dense model as the “#3 open model in the world” and the 26B MoE model at #6 on Arena AI’s text leaderboard, as of that post.

Where the Gemini line is Google’s proprietary, API-only family, Gemma 4 is the family you download. Four sizes shipped at launch; there are five today. Read any “four models” and “five models” description of Gemma 4 as a difference in date, not a contradiction — the 12B Unified model was added on June 3, 2026.

The other number worth reading carefully is the name. E2B and E4B are not 2B and 4B models. The “E” stands for effective parameters — the footprint activated during inference — and the checkpoints are roughly twice that size on disk. See Technical Specifications for both figures side by side.

Release Timeline

The April 2 date is the announcement, not a separate model: the weights were already on Hugging Face and Kaggle. The Android AICore Developer Preview on the same day is an Android-specific distribution channel — Gemma 4 is the foundation for the next generation of Gemini Nano, and code written against the preview is meant to carry over to Gemini Nano 4 devices. The QAT release is dated from Google’s own post; it does not appear as a line in the release-notes table.

Model Family

Gemma 4 E2B

2.3B effective parameters, 5.1B with embeddings. Mobile- and browser-optimised, multimodal, with native audio input. 128K context window.

Gemma 4 E4B

4.5B effective parameters, 8B with embeddings. Targets edge devices. Multimodal with native audio. 128K context window.

Gemma 4 12B Unified

Released June 3, 2026. An 11.95B parameter encoder-free multimodal model: vision and audio encoders are replaced by direct projections of the input into the token space, so the LLM backbone processes every modality natively. Google’s first mid-sized Gemma with native audio input, and small enough to run with 16GB of VRAM or unified memory. Up to 256K context.

Gemma 4 26B A4B

A Mixture-of-Experts model: 25.2B total parameters with 3.8B active per token. Prioritises throughput — Google positions it as “a highly efficient 26B MoE model designed for high-throughput, advanced reasoning,” and as the faster choice against the dense 31B. Text and image input only. Up to 256K context.

Gemma 4 31B

Dense, 30.7B parameters, and the quality ceiling of the family. Text and image input only. Up to 256K context.

Capabilities

• Advanced reasoning and agentic workflows: Google’s stated design target for the generation.
• Native multimodality: Text and image input across all five variants, with variable aspect ratio and resolution support; native audio and video on E2B, E4B and 12B Unified.
• Encoder-free multimodal processing (12B Unified): raw audio projected straight into the token space, with a lightweight embedding module standing in for the vision encoder.
• 140+ languages: natively trained rather than adapted after the fact.
• Long context: 128K on the edge models, up to 256K on the 12B, 26B A4B and 31B.
• Local deployment: 16GB of VRAM or unified memory is enough for the 12B Unified model; the E-series targets phones and browsers, and Google’s Android figures claim up to 4x the speed of the previous generation at up to 60% less battery.
• Multi-token prediction: MTP checkpoints of the four launch sizes for faster decoding.
• Quantization-aware training: 4-bit checkpoints trained to tolerate quantization, plus a mobile format, across the family.

Technical Specifications

The family, in full

What “Effective” means

Google’s naming for the two edge models describes what they activate, not what they store. The launch post says E2B and E4B “activate an effective 2 billion and 4 billion parameter footprint during inference”; the model card puts the exact figures at 2.3B and 4.5B effective, against 5.1B and 8B total once embeddings are counted. Google’s own QAT write-up refers to deploying text-only “without per-layer embeddings,” which is the mechanism the gap comes from.

The practical reading, and the reason a “2.3B to 30.7B” range and a “2B to 31B” range are both honest descriptions of the same family:

• Size the disk and the download against the total. E2B is a 5.1B-parameter file, not a 2.3B one.
• Size latency, battery and activated memory against the effective count. That is the number the mobile deployment story rests on.
• The 26B A4B has the same split for a different reason. Its 3.8B active per token is MoE routing, not offloaded embeddings — 25.2B parameters still have to be resident.
• The 31B and 12B have no gap. Dense models activate what they store.

Quantization-aware training (June 5, 2026)

Google published QAT checkpoints for the family on June 5, 2026. Quantization is simulated during training, so the weights learn to absorb the rounding error rather than meeting it cold at export — Google reports quality above standard post-training quantization at the same bit width. See quantization for the underlying trade-off.

Two formats shipped: a Q4_0 GGUF for general use, and a mobile-specific format using a custom wNa8o8 schema with 2-bit token-generation layers, which Google says brings the E2B footprint down to 1GB — under 1GB for text-only deployment without per-layer embeddings. A compressed-tensors -w4a16-ct serialisation is also offered for high-concurrency cloud serving.

Everything else

• License: Apache 2.0, verbatim
• Languages: 140+, natively trained
• Input modalities: text and image (all variants); audio and video (E2B, E4B, 12B Unified)
• Output modalities: text
• Audio limit: 30 seconds maximum
• Video limit: 60 seconds maximum, processed at one frame per second
• Knowledge cutoff: January 2025
• Checkpoint variants: base, multi-token prediction (E2B, E4B, 26B A4B, 31B), and QAT

Use Cases

• On-device assistants: E2B and E4B on phones, browsers and wearables, with native audio input for voice interfaces and a QAT mobile build that fits in about 1GB.
• Laptop-local multimodal agents: 12B Unified handles voice, images and text on a machine with 16GB of VRAM or unified memory, with no cloud round-trip.
• Latency-sensitive serving: 26B A4B activates only 3.8B parameters per token, which Google positions as the high-throughput option against the dense 31B.
• Fine-tuning base: Apache 2.0 weights across five sizes, so the base checkpoint can be matched to the deployment target rather than the other way round.
• Private and air-gapped deployment: open weights, no API dependency, no per-token cost.
• Long-document work: 256K context on the mid and large variants for codebases and document archives.
• Multilingual products: 140+ languages trained natively rather than bolted on.

Performance / Benchmarks

Google publishes a full benchmark table in the Gemma 4 model card. These vendor-reported numbers are the two ends of the family — they show what the effective parameters of an edge model buy, and what they cost:

Note the exact benchmark names: these are MMLU Pro and GPQA Diamond, the harder variants. Google reports no plain MMLU and no HumanEval score for Gemma 4. The model card carries additional rows and covers the E4B, 12B and 26B A4B variants as well.

Separately, in its April 2, 2026 launch post Google reported placements on Arena AI’s text leaderboard, a third-party community board:

• Gemma 4 31B — “#3 open model in the world”
• Gemma 4 26B A4B — #6

Both are open-model rankings, not overall standings, and both are as of the launch post. Google’s summary claim, on the same evidence, is that Gemma 4 outperforms “models 20x its size.” Community leaderboards move as new models are submitted, so check the live board before relying on a position that is now five months old.

Limitations

• Knowledge cutoff of January 2025. Roughly twenty months stale as of September 2026; pair with retrieval for anything current.
• Text output only. The models understand audio, video and images; they do not generate them.
• Audio and video are not family-wide. The 26B A4B and 31B accept text and images only — their model-card modality rows read “Text, Image.” If you need audio or video from an open-weight Google model, the ceiling is the 12B, not the 31B.
• Video is frames, and short. 60 seconds maximum at one frame per second, so fast motion and audio-visual sync are outside what the model sees. Audio caps at 30 seconds.
• Effective is not actual. E2B is 2.3B effective but 5.1B with embeddings, and E4B is 4.5B against 8B. The on-disk footprint is roughly double what the name suggests — a recurring source of failed deployments to devices sized against the label.
• Hardware floor still applies. 12B Unified needs 16GB of VRAM or unified memory at full precision; the 31B dense model needs substantially more. QAT lowers this but does not remove it.
• Open weights, real responsibility. Apache 2.0 grants broad rights but removes the vendor-side safety layer a hosted API provides.
• Vendor-run benchmarks. Every figure in the table above is Google’s own, on Google’s harness. Treat them as a starting point for your own evaluation.

Pricing & Access

Gemma 4 weights are free to download under Apache 2.0 and free to run on your own hardware. There is no per-token price from Google for self-hosted use.

Where to get the weights
– Hugging Face — base, MTP and QAT collections
– Kaggle
– Ollama
– LM Studio

Hosted options
– Google AI Studio and the Gemini API
– Google Cloud deployment (Vertex AI, Cloud Run)
– Android AICore, for on-device inference in Android apps
– Third-party inference providers

Commercial use is permitted under the Gemma 4 license, which is verbatim Apache License 2.0. The older Gemma Terms of Use — which bind users to a Prohibited Use Policy — explicitly do not govern Gemma 4; that page now redirects Gemma 4 users to the Apache 2.0 license.

Ecosystem & Tools

Local runtimes

• Ollama – One-command local serving
• LM Studio – Desktop GUI for quantized checkpoints
• llama.cpp / GGUF – CPU and Apple Silicon inference; the QAT Q4_0 checkpoints target this path

Frameworks

• Hugging Face Transformers – Reference implementations
• JAX, PyTorch, Keras – Official model definitions
• Google Cloud – Vertex AI and Cloud Run deployment paths
• Android AICore – On-device serving, and the bridge to Gemini Nano 4

Related

• Gemini 3.8 Flash – Google’s proprietary API model line
• Llama 4 – The other major open-weight multimodal MoE family

Community & Resources

• Gemma 4 launch post – April 2, 2026
• Introducing Gemma 4 12B – June 3, 2026, the encoder-free model
• Gemma 4 with quantization-aware training – June 5, 2026, the QAT checkpoints
• Announcing Gemma 4 in the AICore Developer Preview – April 2, 2026, the Android rollout
• Gemma release notes – Exact dates for every checkpoint
• Gemma 4 model card – Parameter counts, modalities, knowledge cutoff, full benchmark table
• Gemma core model docs – Variants, context windows, modalities, QAT formats
• Gemma portal
• Gemma 4 license (Apache 2.0)
• Hugging Face collection