gemma-4-31B-it

by google

Top-3 open VLM with 256K context for dense visual document understanding

820Kdl/month

31Bparams

HuggingFace Use in Pipeline

Identifiers

Model ID

google/gemma-4-31B-it

Feature URI

mixpeek://image_extractor@v1/google_gemma4_31b_v1

Overview

Gemma 4 31B is Google's dense vision-language model, currently ranked #3 among open models on the Arena AI text leaderboard. Unlike the MoE variant (27B-A4B), this dense model activates all 31B parameters, delivering the highest quality at higher compute cost.

The 256K context window and built-in thinking mode make it particularly strong for complex document understanding tasks where accuracy matters more than throughput.

Architecture

Dense transformer architecture with 31B parameters. Vision encoder processes image patches. 256K context window. Thinking mode enables chain-of-thought reasoning for complex visual tasks.

Mixpeek SDK Integration

from mixpeek import Mixpeek
mx = Mixpeek(api_key="YOUR_KEY")

mx.ingest(
    collection_id="technical-docs",
    source="s3://diagrams/",
    extractors=[
        {
            "type": "scene_caption",
            "model": "google/gemma-4-31B-it",
            "output_feature": "caption"
        },
        {
            "type": "text_embedding",
            "model": "Qwen/Qwen3-Embedding-4B",
            "input_field": "caption",
            "output_feature": "caption_embedding"
        }
    ]
)