Gemma 4 31B

Deploy Gemma 4 31B on a dedicated GPU from $0.45 per hour, billed per minute, on 1× RTX 4090. Google's 30.7B dense Gemma 4 model for text and vision, combining a 256K context window with configurable thinking, native function calling and strong reasoning and coding scores.

Multimodal30.7B paramsApache 2.0Deploys in ~5 minSecure Cloud

google/gemma-4-31B-it

HexGrid Cloud price

$0.45/hr

1× RTX 4090

Context

256K

tokens

Parameters

30.7B

dense

Input

Text + Image

multimodal

Running Gemma 4 31B on HexGrid Cloud

What you get when you deploy with us, beyond the hourly rate.

Single tenant by default

The model runs on a GPU that is yours for the life of the instance. No shared endpoint, no queue behind other tenants.

Per-minute billing

You pay for the GPU, not the token. Stop the instance and billing stops with it.

OpenAI-compatible endpoint

Point an existing SDK at your instance by changing the base URL. No rewrite required.

Your weights, your data

Prompts and outputs stay inside your instance. Nothing is logged, sampled or used for training.

Gemma 4 31B GPU sizing and cost

What the model needs at each precision, and the cheapest HexGrid Cloud configuration that holds it.

Gemma 4 31B vRAM requirement and hourly cost by precision
Precision	vRAM needed	Runs on	Price
BF16 Recommended	74 GB	5× RTX A4000	$1.00/hr
FP8	37 GB	1× RTX A6000	$0.55/hr
INT4	19 GB	1× RTX 4090	$0.45/hr

No hardware-independent throughput number is published. Benchmark the precise HexGrid GPU, context length, image resolution and batch configuration before publishing tok/s.

About Gemma 4 31B

What Gemma 4 31B is built for, and where it falls short.

Gemma 4 31B is Google's largest dense Gemma 4 checkpoint. The instruction-tuned version accepts interleaved text and images and generates text, with a 256K-token context window.

Its architecture uses 60 decoder layers, a hybrid of local sliding-window and global attention, and a dedicated approximately 550M-parameter vision encoder.

Gemma 4 adds configurable thinking, native structured function calling, stronger coding capabilities and support for video understanding through sequences of image frames.

What people run it for

Multimodal assistants

Assistants that can reason jointly over screenshots, photographs, diagrams and text.

Document understanding

Extract, summarize and reason over visually structured documents and long textual context.

Coding agents

Generate and modify code while using structured functions or external development tools.

Visual question answering

Answer questions about diagrams, UI screenshots, charts and real-world images.

Strengths

256K context for large documents, long agent histories and multimodal prompts
Native image understanding plus interleaved text-image input
Configurable thinking mode for trading reasoning depth against latency
Native function calling for agentic applications
Apache 2.0 licence with straightforward commercial deployment terms
Pre-trained on material spanning more than 140 languages

Limitations

The 31B checkpoint does not include the audio encoder available in smaller Gemma 4 variants
30.7B dense parameters make full-precision deployment substantially heavier than smaller Gemma variants
Video is handled through sampled image frames rather than a dedicated video encoder
256K-context serving can require substantial KV-cache memory
Benchmark results depend on Google's documented evaluation settings and should not be treated as workload-specific guarantees

Quickstart

Deploy behind an OpenAI-compatible multimodal chat endpoint and submit text or image-plus-text messages.

pythonPOST /v1/chat/completions

from openai import OpenAI

client = OpenAI(
    base_url="https://<your-instance>.hexgrid.cloud/v1",
    api_key="<your-key>",
)

response = client.chat.completions.create(
    model="google/gemma-4-31B-it",
    messages=[
        {"role": "user", "content": "Explain the tradeoff between local and global attention."}
    ],
)

print(response.choices[0].message.content)

Gemma 4 is natively multimodal. Image requests should use the multimodal content-array format supported by the serving engine.

Gemma 4 31B reported benchmarks

Scores published by the model authors. A starting point, not a guarantee on your own data.

MMLU Pro85.2%
AIME 2026 no tools89.2%
LiveCodeBench v680%
GPQA Diamond84.3%
Codeforces2150 ELO
MMMU-Pro76.9%

Gemma 4 31B specifications

Architecture and serving details for Gemma 4 31B.

PublisherGoogle DeepMindParameters30.7BLicenceApache 2.0ReleasedApril 2026

Architecture

Type: Dense multimodal Transformer
Parameters: 30.7B
Layers: 60
Sliding window: 1024 tokens
Vision encoder: ~550M parameters
Vocabulary: 262K

Context and modalities

Context window: 256K tokens
Input: Text + image
Output: Text
Audio: Not supported on 31B
Video: Supported through image-frame sequences

Capabilities

Thinking mode: Configurable
Function calling: Native
Languages: 35+ out of the box; pretrained on 140+
Licence: Apache 2.0

Gemma 4 31B frequently asked questions

Common questions about deploying Gemma 4 31B on HexGrid Cloud.

Is Gemma 4 31B actually a 31-billion-parameter model?

Google reports 30.7B total parameters; 31B is the rounded family name.

Does Gemma 4 31B support audio?

No. Google's table lists audio for E2B, E4B and 12B, but not for the 31B model.

Does it support images?

Yes. Gemma 4 31B has a dedicated approximately 550M-parameter vision encoder and supports interleaved text-image input.

Can I enable or disable reasoning?

Yes. Google's model interface exposes configurable thinking behavior.

Can I use it commercially?

Yes. Gemma 4 is released under Apache 2.0.

Deploy Gemma 4 31B today

$0.45 per hour on 1× RTX 4090, billed per minute, never shared.

Deploy Gemma 4 31B See all models