Gemma 4

Overview

Gemma 4 is Google DeepMind's current open-weight model family, first released on March 31, 2026 and expanded through June 2026. It is licensed under Apache 2.0 — genuinely permissive, not a bespoke community license — with open weights and responsible commercial use permitted.

Google's framing is intelligence density: the family is "purpose-built for advanced reasoning and agentic workflows," and the company's headline claim is that "Gemma 4 outcompetes models 20x its size." The evidence it offers is leaderboard placement — the 31B dense model at #3 and the 26B MoE model at #6 among open models on the Arena text leaderboard.

Where Gemini 3.5 is the proprietary API model, Gemma 4 is the family you download. The five variants stretch from a 2.3B-effective-parameter mobile model to a 30.7B dense flagship. Every one of them handles text, images, and video-as-frames; native audio input is reserved for E2B, E4B, and the 12B Unified model.

Release Timeline

Model Family

Gemma 4 E2B

2.3B effective parameters (5.1B with embeddings). Mobile-optimized, 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

The newest member, released June 3, 2026. An 11.95B parameter encoder-free multimodal model: vision and audio encoders are replaced by direct linear 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. Prioritizes latency — Google's model card notes it "runs much faster than its 26B total might suggest," making it "an excellent choice for fast inference compared to the dense 31B model." Ranks #6 among open models on the Arena text leaderboard per Google. Up to 256K context.

Gemma 4 31B

Dense, 30.7B parameters, and the quality ceiling of the family. Ranks #3 among open models on the Arena text leaderboard per Google. 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; native audio on E2B, E4B, and 12B Unified.
• Video understanding: All variants analyze video by processing sequences of frames — a maximum of 60 seconds at one frame per second.
• 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 across more than 140 languages.
• Long context: 128K on edge models, up to 256K on the 12B, 26B, and 31B.
• Local deployment: 16GB of VRAM or unified memory is enough for the 12B Unified model; the E-series targets phones.
• Multi-token prediction: MTP checkpoints of E2B, E4B, 26B A4B, and 31B for faster decoding.

Technical Specifications

Every variant accepts text and images, and every variant handles video as a frame sequence. The audio column is the only modality that splits the family.

• License: Apache 2.0 (the "Gemma 4 license")
• Languages: 140+, natively trained
• Input modalities: Text, image, video-as-frames (all variants); audio (E2B, E4B, 12B Unified)
• Output modalities: Text
• Video limit: 60 seconds maximum, processed at one frame per second
• Knowledge cutoff: January 2025
• Variants: Base checkpoints plus multi-token prediction (MTP) checkpoints for E2B, E4B, 26B A4B, and 31B

Use Cases

• On-device assistants: E2B and E4B on phones and wearables, with native audio input for voice interfaces.
• 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 "an excellent choice for fast inference compared to the dense 31B model."
• Fine-tuning base: Apache 2.0 weights across five sizes, so the base checkpoint can be matched to the deployment target.
• Private and air-gapped deployment: Apache 2.0 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 are the two ends of the family — the numbers show what the "effective parameters" of an edge model actually buy you, 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 variants as well.

Separately, Google reports placements on the Arena text leaderboard, a third-party community leaderboard:

• Gemma 4 31B — #3 among open models
• Gemma 4 26B A4B — #6 among open models

Both are open-model rankings, not overall standings. Google's summary claim is that "Gemma 4 outcompetes models 20x its size." Leaderboard rankings shift as new models are submitted, so check the live board before relying on a specific position.

Limitations

• Knowledge cutoff of January 2025. Roughly eighteen months stale as of mid-2026; pair with retrieval for current information.
• Text output only. The models understand audio, video, and images; they do not generate them.
• Native audio input is not universal. The 26B and 31B accept text, images, and video frames — but not audio.
• Video is frames, not video. All variants cap out at 60 seconds sampled at one frame per second, so fast motion and audio-visual sync are outside what the model sees.
• Hardware floor still applies. 12B Unified needs 16GB of VRAM or unified memory; the 31B dense model needs substantially more.
• Effective vs. actual parameters. 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.
• Open weights, real responsibility. Apache 2.0 grants broad rights but removes the vendor-side safety layer that hosted APIs provide.

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
• Kaggle
• Ollama
• LM Studio

Hosted options

• Google AI Studio and the Gemini API
• Google Cloud deployment (Vertex AI, Cloud Run)
• Third-party inference providers

Commercial use is permitted under the Gemma 4 license, which is verbatim Apache License 2.0. Note that 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

Frameworks

• Hugging Face Transformers - Reference implementations
• JAX, PyTorch, Keras - Official model definitions
• Google Cloud - Vertex AI and Cloud Run deployment paths

Related

• Gemini 3.5 - 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 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
• Gemma portal
• Gemma 4 license (Apache 2.0)
• Hugging Face collection