JA EN
shiichan

Beating Models 20x Its Size — Gemma 4 31B Lands on SageMaker JumpStart!

AWS What's New ★★★★ View Markdown
#aws#ai

Hi everyone, I'm Shiichan! Today's news is about Google DeepMind's newest model, Gemma 4, arriving on SageMaker JumpStart — in both a full-power version and a lightweight one.

AWS What's New aws.amazon.com

What was announced?

AWS What's New announced that two models are now available on Amazon SageMaker JumpStart: Google DeepMind's Gemma-4-31B-it-assistant, and NVIDIA's quantized version, Gemma-4-31B-IT-NVFP4. Both are built on the same Gemma 4 31B (31 billion parameter) dense model — one at full precision, one optimized and quantized.

Why it matters

Bigger models tend to perform better, but they also cost more to run in memory and compute. Here, the same base model comes in both a "use it as-is" version and a "lightened" version, so you can choose based on whether your workload prioritizes accuracy or cost and speed.

What changes

Here's what each model brings:

  • Gemma-4-31B-it-assistant: An assistant-tuned model built for multimodal reasoning, coding, and agentic workflows. It handles text and image input (with video treated as frame sequences), supports a 256K-token context window, and covers over 140 languages. It combines local sliding-window attention with global attention in a hybrid mechanism, plus native function calling for building autonomous agents. On the Arena AI text leaderboard, it ranked #3 — beating models 20 times its size.
  • Gemma-4-31B-IT-NVFP4: The same Gemma 4 31B capabilities, at a much smaller memory footprint. NVIDIA's ModelOpt framework quantizes it to 4-bit FP4 precision, cutting memory usage to about 18.5 GB (68% smaller than the base model) and delivering roughly 2.5x faster inference, while retaining 97-99% of the original model's quality.

Dive Deep

Gemma-4-31B-IT-NVFP4 is optimized for NVIDIA RTX, DGX Spark, and data center GPUs. If your team's GPU resources are limited, this quantized version lets you get close to Gemma 4 31B-class performance with a much smaller memory budget. Deploy either model from the SageMaker JumpStart model catalog in a few clicks, or via the SageMaker Python SDK.

Wrap-up

  • Gemma-4-31B-it-assistant (full precision) and Gemma-4-31B-IT-NVFP4 (quantized) are now on SageMaker JumpStart
  • The full-precision version supports 256K tokens, 140+ languages, and beats models 20x its size on the Arena AI leaderboard
  • The quantized version cuts memory by 68% and runs 2.5x faster, while keeping 97-99% of the quality
  • Pick the full-precision version for accuracy, or the quantized version for cost and speed, depending on your workload

Related posts