Gemma (Google)

Gemma 4 needs 5GB VRAM (E2B/E4B), 16GB (26B MoE), or 24-62GB (31B Dense) depending on quantization. Requirements by model: RTX 3060 to H100, Apple Silicon M1-M4, CPU-only operation, RAM sizing, and budget builds. Updated July 2026.

Comprehensive comparison of Gemma 4, Llama 4, and Qwen 3.5 local LLMs. Detailed analysis of benchmark performance, licensing, Japanese support, hardware requirements, and use case selection criteria.

Complete guide to Google Gemma 4 (released April 2, 2026): 4 model variants (E2B/E4B/26B MoE/31B Dense), Apache 2.0 license, system requirements, multimodal capabilities, AIME 89% benchmark, 140+ languages, and step-by-step Ollama installation and setup instructions.

**On 2026-06-29 Cerebras launched [Gemma 4 31B on Cerebras Inference](https://www.cerebras.ai/blog/gemma-4-on-cerebras-the-fastest-inference-is-now-multimodal)** — the platform's **first multimodal model** and the **first Google DeepMind model** available on it, in public preview. **Performance**: **1,851 output tokens/second** (**35× a typical GPU endpoint**), **≤1.5 s to first token including reasoning**, **[Artificial Analysis Intelligence Index of 29](https://artificialanalysis.ai/)** (comparable to Claude Haiku 4.5's 30), and **18× the speed of Haiku on Cerebras**. **Model specs**: **[Gemma 4 31B](../columns/gemma-4-technical-report-2026-07) dense architecture** (not MoE), **Apache 2.0 open weights**, long-context capable, and image understanding (screenshots, charts, documents, UI states, diagrams, scanned pages, forms). **Distinctive value**: unlocks multimodal workloads that used to be impractical on GPUs — **computer use, image-driven agentic loops, UI debugging with code-patch generation, real-time dashboard analysis, long-document summarization** — at wafer-scale speed. **Position**: the [Gemma 4 Technical Report](../columns/gemma-4-technical-report-2026-07) covered Google DeepMind's open-weights strategy; combined with **Cerebras's wafer-scale hardware, it reaches a new "multimodal × fast inference × open weights" infrastructure stack** — the open-weights + fast-inference camp's counter to the closed-model camp underlying [OpenAI GPT-5.6 + ChatGPT Work](../columns/openai-chatgpt-work-launch-2026-07) and [Claude Cowork](../columns/claude-cowork-web-mobile-launch-2026-07). Paired with [Nous Portal's 300+ neutral models](../columns/nous-portal-cloud-hermes-agent-2026-07) or [local LLM deployments](../columns/local-llm-landscape-2026-june-update), it becomes part of the **late-2026 open-weights-practical AI infrastructure stack**. **Cerebras Inference Cloud public preview** (limited time), with details in the [API docs](https://inference-docs.cerebras.ai/models/gemma-4-31b).

**[Software Mansion's React Native ExecuTorch](https://docs.swmansion.com/react-native-executorch/) shipped a `useLLM` hook** that gives React Native apps **native on-device LLM integration**. **Supported models**: quantized Qwen (2.5 / 3 / 3.5), Llama 3.2, Hammer 2.1, Phi 4 Mini, SmolLM 2, LFM2.5 (vision), and **[Gemma 4](../columns/gemma-4-technical-report-2026-07) (vision + audio)**. **Two operating modes**: **Functional/Stateless** (developers manage conversation history via `generate()` + `response`; tool calling and chat config don't apply) and **Managed/Stateful** (`sendMessage()` maintains conversation state, parses tool calls, and runs callbacks automatically). **Key features**: token batching (groups tokens before re-render), tool calling (model invokes external functions via tool schemas with automatic parsing and callbacks), vision-language / audio multimodal inputs, generation control (temperature / top-p / repetition penalty / mid-stream interruption), and **JSONSchema- or Zod-backed structured output**. **Use cases**: on-device chatbots without server dependencies, privacy-first conversation UIs, in-app function calling (calendar events, flashlight, etc.), multimodal features (image analysis, audio transcription). **Positioning**: implementation-level evidence that **cutting-edge open-weight LLMs like [Gemma 4's encoder-free 12B](../columns/gemma-4-technical-report-2026-07) and [Qwen 3.6-35B](../columns/qwen36-35b-a3b-uncensored-abliterated-2026-07) natively fit into iOS / Android apps** — a major deliverable from the React Native ecosystem.

**Google DeepMind's Gemma Team released the Gemma 4 Technical Report as [arXiv:2607.02770](https://arxiv.org/abs/2607.02770) on 2026-07-02**. The paper introduces **2.3B / 12B / 31B parameter models**, **both Dense and MoE variants**, **natively multimodal (text / image / audio)**, a **12B encoder-free unified design** (raw audio and image patches processed directly without separate encoders), a **built-in reasoning (thinking) mode**, **improved vision / audio encoders**, **architectural refinements for inference speed, memory efficiency, and long context**, and **competitive performance against larger open models on STEM, multimodal, and long-context benchmarks**. Over 300 authors contributed. Open weights allow commercial use, distributed via Hugging Face and Ollama. Sits alongside [Qwen 3.6-35B-A3B](../columns/qwen36-35b-a3b-uncensored-abliterated-2026-07) and the [Local LLM June 2026 update](../columns/local-llm-landscape-2026-june-update) as a new chapter at the open-weights frontier. **A milestone in Google's open-weights strategy**; the encoder-free unified design departs from Qwen / Llama / DeepSeek multimodality (separate vision encoder + projection). The **reasoning mode** mirrors the extended-thinking modes of Anthropic and OpenAI closed models — the open ecosystem catching up. Caveats: commercial-license fine print, potential systemic-risk classification (EU AI Act's 10^25 FLOPs threshold), and heavy Google Cloud Vertex AI integration bias.