[AINews] Gemma 4: The best small Multimodal Open Models, dramatically better than Gemma 3 in every way

2026-04-03 · Source: Latent.Space - Www.latent.space · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, medium

Summary

Google DeepMind has released Gemma 4, a family of open-weight models under the commercially permissive Apache 2.0 license, marking a significant advancement in open-source AI. The 31B dense variant ties with Kimi K2.5 (744B-A40B) and Z.ai GLM-5 (1T-A32B) as a top open model, but with substantially fewer parameters. Gemma 4 models are designed for reasoning and agentic workflows, supporting local/edge deployment, and feature native multimodal capabilities for video, images, and audio, including OCR, chart understanding, and speech recognition. Key architectural choices include hybrid attention mechanisms, proportional RoPE, Per-Layer Embeddings (PLE), KV-cache sharing, and long context up to 256K. The release saw immediate ecosystem support across platforms like llama.cpp, Ollama, and vLLM, with impressive local inference performance demonstrations.

Key takeaway

For CTOs and VP of Engineering evaluating open-source AI models for multimodal applications or edge deployment, Gemma 4 represents a compelling option. Its Apache 2.0 license, efficient architecture, and native support for video, image, and audio processing, combined with strong benchmark performance, make it suitable for developing reasoning and agentic workflows. You should explore its integration into your existing infrastructure, especially given its immediate support across popular local inference frameworks, to capitalize on its capabilities for on-device or resource-constrained environments.

Key insights

Gemma 4 advances open-source AI with multimodal capabilities, efficient architecture, and a permissive Apache 2.0 license.

Principles

• Smaller models can achieve top-tier performance.
• Multimodal input is crucial for edge AI.
• Open-source licensing drives rapid adoption.

Method

Gemma 4 employs a hybrid 5:1 local/global attention mechanism, partial-dimension RoPE, and per-layer embeddings, alongside a refined training recipe and data, to achieve high performance with fewer parameters.

In practice

• Deploy Gemma 4 for local/edge multimodal AI tasks.
• Utilize its function calling for agentic workflows.
• Integrate with existing stacks via day-0 support.

Topics

• Gemma 4
• Multimodal AI Models
• Open-weight Licensing
• AI Agent Workflows
• Edge AI Deployment

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Engineer, Machine Learning Engineer, AI Scientist

Related on AIssential

• Gemma 4 — Gemma 4 offers powerful, open-weight multimodal AI with strong local deployment and agentic capabilities under an Apache 2.0 license.
• Welcome Gemma 4: Frontier multimodal intelligence on device — Gemma 4 offers open, multimodal AI models optimized for on-device deployment and broad ecosystem integration.
• Gemma 4 Local Test | New Open LLM King? — Gemma 4 offers strong multimodal capabilities and agentic workflow support under an Apache 2 license.
• Google’s Gemma 4: Is it the Best Open-Source Model of 2026? — Google's Gemma 4 family offers versatile, open-weight LLMs optimized for diverse environments, from edge devices to consumer GPUs.
• Gemma 4: Byte for byte, the most capable open models — Gemma 4 offers highly capable, open-source AI models optimized for diverse hardware, from edge devices to data centers.
• Google just dropped Gemma 4... (WOAH) — Gemma 4 offers high intelligence per parameter, enabling powerful AI on consumer and edge devices.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Latent.Space - Www.latent.space.