Google DeepMind and NVIDIA Offer Gemma-4-31B Models for Multimodal AI and Cost-Efficient Inference

AWS added Gemma-4-31B dense models—full-precision and NVIDIA-quantized versions—providing large-scale multimodal capabilities with optimized efficiency.

By Lydia Stone · Editorial pseudonym · 2026-09-14 · 1 min read

AI-ASSISTED REPORTING · RETROSPECTIVE ARCHIVE

Retrospective edition: 2026-09-14. Source published: 2026-09-14. Added to this archive: 2026-11-03. This briefing was prepared later from the linked announcement; it was not published on this website on the historical edition date.

Gemma-4-31B-it-assistant is a multimodal model supporting text and image inputs with a large 256K-token context window and over 140 language support, designed for agentic workflows and autonomous agents. It integrates hybrid attention mechanisms enabling efficient reasoning and native function calling.

The Gemma-4-31B-IT-NVFP4 variant is quantized to 4-bit FP4 precision using NVIDIA ModelOpt, significantly reducing memory footprint and accelerating inference by approximately 2.5x while retaining near-original quality. These models enable scalable AI deployments on NVIDIA GPUs via SageMaker JumpStart.

Source & further reading

Gemma-4-31B-it-assistant and Gemma-4-31B-IT-NVFP4 models now available on Amazon SageMaker JumpStart ↗

Source published 2026-09-14. This report was prepared with AI assistance from the linked primary-source material. Claims and limitations should be checked against the original announcement or research.