Gemma 4: Frontier Multimodal Intelligence You Can Run Anywhere

Gemma 4 introduces native multimodal support (text, image, video, audio), extended context windows (128K-256K), configurable thinking modes, and built-in function calling. The architecture uses Per-Layer Embeddings (PLE) for efficiency and shared KV cache to reduce memory usage during long-context generation.

E2B (2.3B effective) and E4B (4.5B effective) are designed for edge devices, browsers, and mobile with native audio support. The 26B A4B is a Mixture-of-Experts model activating only 4B parameters per token, ideal for high-throughput scenarios. The 31B dense model is the flagship for maximum performance on reasoning, coding, and vision tasks.

All models process text and images with variable aspect ratios natively. The vision encoder uses learned 2D positions and can encode images to different token budgets (70-1120 tokens) for speed-quality tradeoffs. E2B and E4B include USM-style conformer audio encoders for speech understanding. Video is supported across the family by processing frames and audio tracks.

PLE gives each decoder layer its own small embedding for every token, creating a parallel conditioning pathway alongside the main residual stream. This allows each layer to receive token-specific information only when relevant, rather than packing everything into a single upfront embedding. It adds meaningful per-layer specialization at modest parameter cost, making small models more efficient.

Gemma 4 models are available on Kaggle and Hugging Face under Apache 2.0 license. You can use them through Google AI Studio, deploy on Vertex AI, or run locally with tools like Ollama, llama.cpp, MLX (for Apple Silicon), transformers, and Mistral.rs. ONNX checkpoints enable browser and edge device deployment.

E2B requires ~9.6GB (BF16) to 3.2GB (4-bit) VRAM. E4B needs ~15GB (BF16) to 5GB (4-bit). The 31B model requires ~58GB (BF16) to 17GB (4-bit). The 26B MoE needs ~48GB (BF16) to 16GB (4-bit). These are base weights only; add memory for context window (KV cache) based on your use case.

Yes. The E2B and E4B models are specifically designed for browser and mobile deployment. transformers.js enables running Gemma 4 directly in browsers with WebGPU support. ONNX checkpoints work on various edge hardware backends. The models are optimized for Pixel devices and Chrome browser environments.

Gemma 4 has built-in function calling support without requiring fine-tuning. The models can parse tool definitions, generate structured JSON calls, and handle multimodal function calling (e.g., analyzing an image and calling a weather API). This powers autonomous agents for tasks like code execution, web browsing, and data retrieval.

The 31B model ranks #3 on Arena AI leaderboard among open models, ahead of Llama 3.3 70B despite being less than half the size. It achieves 89.2% on AIME 2026 math reasoning, 85.2% on MMLU Pro, and 80% on LiveCodeBench v6. The efficiency comes from architectural innovations like alternating attention patterns and shared KV cache.

The 26B A4B model has 26 billion total parameters but activates only 4 billion per token during generation. All 26B parameters must be loaded into memory for fast routing, but inference cost is closer to a 4B model. This achieves 88.3% on AIME 2026 and 82.6% on MMLU Pro with significantly lower compute per token than the dense 31B model.

Yes. Small models support 128K context windows, while medium models handle 256K tokens. The architecture uses dual RoPE configurations (standard for sliding layers, pruned for global layers) to enable longer context. Shared KV cache reduces memory consumption during long-context generation, making it practical for processing entire codebases and research papers.

Gemma 4 is fully supported in TRL (Transformer Reinforcement Learning), with examples for multimodal tool responses and environment interaction. Hugging Face provides fine-tuning guides for Vertex AI using SFT. Unsloth Studio offers a UI-based fine-tuning experience. The models support PEFT methods like LoRA for parameter-efficient training.