Explore ProductsAI WorkspaceApps & SystemsManaged agentsWhite labelAPI
PlatformsDesktop AppChrome ExtensionWeb App
Explore ResourcesResourcesUse casesCase studies
Mission
SPIRITT logoSPIRITTGoogle × CerebrasGemma 4 31B on Cerebras

Gemma 4 31B on Cerebras runs at 1,800+ tokens per second

Artificial Analysis measures Google's open-weight multimodal Gemma 4 31B at 1,808 output tokens per second on Cerebras, with 0.57s first-token latency and roughly 9x the next-fastest provider. On SPIRITT, that speed powers real coding, visual, and agentic work inside a fully equipped cloud workspace.

Try Gemma 4 free
See how it works

Gemma 4 on Cerebras, seen in the wild

People are stunned by how fast it feels, then start inventing new use cases that only work when the model can keep up.

Official launch card

Gemma 4 31B is Google's flagship dense open-weight model: 30.7B parameters, built-in reasoning, text and image input, tool calling, structured output, and a native 256K context window. Cerebras serves it at wafer-scale speed for real-time visual and agentic loops.

What this speed unlocks

Google's model card reports 89.2% on AIME 2026, 80.0% on LiveCodeBench v6, 76.9% on the Tau2 agent benchmark, and 76.9% on MMMU Pro. It also leads its Gemma-family comparisons on long-context retrieval and document understanding.

Cerebras exposes it through an OpenAI-compatible API with image inputs, parallel tool calling, structured outputs, prompt caching, and reasoning controls. Its hosted limit is 131K context on paid plans, with 40K max output; the underlying model supports 256K context.

1,808 tok/s measured0.57s first token31B parametersApache 2.0 weightsMultimodal256K native contextTool calling

AA Intelligence Index

Near frontier
Claude Haiku 4.5
30
Gemma 4 31B
29

Independent Artificial Analysis score. Gemma 4 31B is one point behind Claude Haiku 4.5 while remaining open weight.

Where Gemma 4 leads and where Cerebras changes the experience

Independent provider measurements plus Google's model-card evaluations across coding, reasoning, agentic tool use, vision, documents, and long context.

Throughput
Winner

AA median output tokens per second

Cerebras
1808
SambaNova
201.7
FriendliAI
73
Together AI
44

Higher is better. Independent provider measurement from Artificial Analysis.

Responsiveness
Winner

AA time to first token

Cerebras
0.6
DeepInfra
0.9
Together AI
1.2
FriendliAI
2.1

Seconds, lower is better. Cerebras has the lowest measured provider latency.

Math reasoning
Winner

AIME 2026, no tools

Gemma 4 31B
89.2%
Gemma 4 26B A4B
88.3%
Gemma 4 12B
77.5%
Gemma 3 27B
20.8%

Google model-card evaluation. Gemma 4 31B leads the family by a wide margin over Gemma 3.

Coding
Winner

LiveCodeBench v6

Gemma 4 31B
80%
Gemma 4 26B A4B
77.1%
Gemma 4 12B
72%
Gemma 3 27B
29.1%

Google model-card evaluation of competitive programming and code generation.

Agentic tool use
Winner

Tau2 average

Gemma 4 31B
76.9%
Gemma 4 12B
69%
Gemma 4 26B A4B
68.2%
Gemma 3 27B
16.2%

Google model-card average over three Tau2 agent environments.

Vision reasoning
Winner

MMMU Pro

Gemma 4 31B
76.9%
Gemma 4 26B A4B
73.8%
Gemma 4 12B
69.1%
Gemma 3 27B
49.7%

Multidiscipline visual reasoning from Google's model card.

Document understanding
Winner

OmniDocBench 1.5 error

Gemma 4 31B
0.1
Gemma 4 26B A4B
0.1
Gemma 4 12B
0.2
Gemma 3 27B
0.4

Average edit distance, lower is better. Google model-card evaluation.

Long context
Winner

MRCR v2, 8 needles at 128K

Gemma 4 31B
66.4%
Gemma 4 26B A4B
44.1%
Gemma 4 12B
43.4%
Gemma 3 27B
13.5%

Long-context retrieval from Google's model card.

Provider tradeoff

AA blended price per 1M tokens

Parasail
$0.14
DeepInfra
$0.15
FriendliAI
$0.17
Cerebras
$1.04

Lower is better. Cerebras wins on speed and latency, not price. Provider pricing changes over time.

Sources: Google DeepMind Gemma 4 model card and technical report; Artificial Analysis Intelligence Index and Gemma 4 provider measurements; Cerebras Gemma 4 launch materials and current inference docs. Provider speed, latency, and price are measured snapshots and can change. Google's capability scores compare Gemma-family models under its published evaluation setup. Cerebras currently exposes 65K context on free tier and 131K paid, below the model's native 256K window.

How It Works

From zero to Gemma 4 31B running real work at Cerebras speed

01

Open a workspace

Start free, including trial. You land in a fully equipped cloud computer: browser, files, terminal, integrations, and memory. No local setup. No thin chat box pretending to be an agent.

Open a SPIRITT workspace for Gemma 4
02

Pick Gemma 4 31B

Open the model picker and choose Gemma 4 31B on Cerebras. The open-weight multimodal model arrives with ultra-fast hosted inference, inside a workspace that already has browser control, files, terminal, and durable context.

Select Gemma 4 31B on Cerebras in the model picker
03

Build or automate

Tell it what to ship or what to run. Gemma 4 can code, call tools, drive the browser, coordinate multi-step work, and keep going while you step away. The point is not another chat window. It is an agentic environment where Gemma 4 actually does the job.

Build at high speed with Gemma 4 on Cerebras

Try Gemma 4 in an agentic environment

A real cloud environment, model picker, tools, and memory. Included in all plans, including trial.

Try Gemma 4 Now
Cerebras inference
Questions

Frequently asked questions

01What is Gemma 4 31B on Cerebras?+
Gemma 4 31B is Google's open-weight multimodal model served on Cerebras' wafer-scale inference infrastructure. Artificial Analysis measures about 1,808 output tokens per second and 0.57s first-token latency on Cerebras, making it unusually fast for coding, visual, and multi-step agent interactions.
02Why does 1,800+ tokens per second matter?+
Agents often call a model repeatedly before finishing one job. Cutting each generation from many seconds to around a second makes coding iterations, browser loops, visual analysis, and voice experiences feel interactive instead of batch-oriented.
03Is Gemma 4 31B open weight?+
Yes. Google describes Gemma 4 31B as an open-weight model under Apache 2.0. Cerebras provides hosted inference, while the weights remain part of the Gemma family rather than a closed proprietary API-only model.
04What are the limitations?+
Speed does not guarantee best-in-class reasoning on every task. On Artificial Analysis, Gemma 4 31B scores near Haiku-class intelligence, not absolute frontier. Cerebras wins on speed and latency, not price. Hosted context is also lower than the model's native 256K window, so test it on your actual workflow.
05What is a good way to test it?+
Give it a visible, iterative job such as building a landing page, analyzing screenshots, or fixing a small application with tests. Those workflows expose whether low latency actually improves the full agent loop, not just a single chat response.
06Where is Gemma 4 31B on Cerebras best to test for full agentic capabilities?+
SPIRITT Workspaces. Pick Gemma 4 31B on Cerebras in the model picker and run real work in a fully equipped cloud environment: tools, browser, files, terminal, memory, and durable sessions. Available on every plan, including trial. That is where the model can act as an agent, not just chat.
Buy from builders who use what they sellBuilt usingSPIRITT