Google goes fully open source with Gemma 4
Google just launched Gemma 4, its most intelligent open model lineup yet, released under a commercially permissive Apache 2.0 license. Spanning four sizes from on-device E2B to a 31B dense flagship, the models handle vision, speech, up to 256K context, and agentic workflows, with its 26B MoE outranking models 20 times its size. It basically delivers frontier-level reasoning to local hardware with minimal overhead.
Gemma 4 is Google's newest lineup of open weights, released under the commercially friendly Apache 2.0 license to bring advanced multi-step reasoning and agentic workflows straight to your local hardware.
The family comes in four sizes: the edge-ready E2B and E4B, alongside larger 26B MoE and 31B Dense variants. Instead of basic chatbots, these models are tuned from the ground up for native function calling, structured JSON outputs, and developer tool execution.
What it can do
Every model in the lineup natively handles interleaved images and video for visual tasks like OCR and chart comprehension. The smaller E2B and E4B variants go a step further with native audio input, enabling offline speech recognition and understanding.
Context windows have expanded significantly. The edge tiers offer 128K tokens, while the 26B and 31B models handle up to 256K tokens, letting you drop entire codebases or dense technical papers into a single prompt. They are pretrained on over 140 languages with broad multilingual support out of the box.
Benchmark performance punches well above its weight. On the Arena AI text leaderboard, the 31B model ranks as the third best open model in the world, while the 26B MoE takes the sixth spot, outcompeting models 20 times its footprint.
The secret is smart parameter efficiency. The 26B MoE activates just 3.8B parameters during inference, delivering rapid token throughput with minimal compute. Meanwhile, the 31B Dense is tuned for raw quality and downstream fine-tuning.
How to run it
Hardware requirements are genuinely accessible. E2B and E4B run completely offline with near-zero latency on Android phones, Raspberry Pi boards, and NVIDIA Jetson Orin Nano modules.
For the larger models, unquantized bfloat16 checkpoints fit onto a single 80GB NVIDIA H100. Quantized versions run smoothly on consumer RTX GPUs, workstations, and DGX Spark setups, pairing easily with tools like OpenClaw for local agent systems.
Ecosystem support is live on day one. Model weights can be grabbed from Hugging Face, Kaggle, or Ollama. You can deploy immediately using Ollama, llama.cpp with GGUF files, LM Studio, vLLM, or MLX, and fine-tune via Unsloth. If you want to skip installation, Google AI Studio hosts the 26B and 31B models, while Google AI Edge Gallery runs the smaller variants.
Why it matters
The big move here is the license. Dropping custom restrictive terms in favor of pure Apache 2.0 removes the cloud hanging over commercial adoption. Developers finally get total control over deployment, infrastructure, and weights.
If you build local agents, coding copilots, or privacy-first document analyzers, Gemma 4 is worth downloading today. The 26B MoE looks like the sweet spot: getting top-tier reasoning while only paying the latency cost of activating 3.8B parameters makes offline local AI truly practical.
Sources (2)
- blog.google(opens in a new tab)Originalblog.google
- blogs.nvidia.com(opens in a new tab)Reportblogs.nvidia.com