Cerebras Runs Gemma 4 at 1,500 Tokens per Second, 15x Faster Than Haiku

Speed has always been Cerebras's pitch. The company built its entire identity around wafer-scale chips , processors the size of an entire silicon wafer , that can run inference at speeds GPU clusters simply cannot match. Until now, that speed advantage was text-only. That changes with the launch of Gemma 4 on Cerebras Inference, which marks the first time the platform supports image inputs, and does so at a pace that redefines what real-time multimodal AI can feel like.

What Just Happened

Gemma 4 is now in private preview on Cerebras Inference, with general availability later this month. It is the first Google DeepMind model Cerebras has brought to the platform, and the first to let developers feed images , screenshots, documents, charts, UI states , into a model running at wafer-scale speed. To celebrate, Cerebras and Google DeepMind are co-hosting a 24-hour hackathon with a $5,000 prize pool, with the top project featured by both companies.

The Number That Matters

Cerebras runs Gemma 4 at over 1,500 output tokens per second. By comparison, Claude Haiku runs at roughly 100 tokens per second , a 15x speedup against the most directly comparable production model, at quality that lands in the same band and at a price lower per output token.

That gap is not just a benchmark flex. Speed compounds in exactly the workloads Gemma 4 is built for. Multimodal and agentic loops rarely call a model once: they inspect a visual input, reason over it, produce structured output, call tools, check the result, and try again. At 100 tokens per second, those loops feel sluggish. At 1,500, they feel instant.

The Model Itself

Gemma 4 31B is the flagship of Google DeepMind's open-weight Gemma family , a dense, multimodal model built for quality and efficiency rather than raw parameter count. Dense models achieve high model intelligence without the large memory footprint of MoE (Mixture-of-Experts) models. MoE is an architecture where only a fraction of the model's parameters activate on any given token, trading memory efficiency for some coordination overhead.

The 31B model currently ranks as the #3 open model in the world on the industry-standard Arena AI text leaderboard, and Gemma 4 outcompetes models 20x its size. On the Artificial Analysis Intelligence Index, Gemma 4 31B scores 29 , effectively matching Claude Haiku at 30, but open-weight under Apache 2.0.

The broader Gemma 4 family is worth understanding:

• Small sizes (E2B, E4B): 2B and 4B effective parameter models built for ultra-mobile, edge, and browser deployment.
• Unified (12B): An encoder-free model for multimodal tasks that replaces vision and audio encoders with direct linear projections of the input.
• Dense (31B): A powerful 31B parameter dense model that bridges the gap between server-grade performance and local execution.
• MoE (26B A4B): A highly efficient 26B MoE model designed for high-throughput, advanced reasoning.

The model Cerebras is running is the 31B dense variant , the one that prioritizes raw intelligence over memory efficiency, which makes sense given that Cerebras's wafer-scale hardware is not memory-constrained in the same way a GPU cluster is.

Why Multimodal Changes the Game for Cerebras

Before this launch, Cerebras's speed advantage was real but narrow: it only applied to text. The moment you needed to process an image, you were back on GPUs. This multimodal model unlocks an entirely new class of applications on Cerebras Inference, from computer use to image-driven agentic workflows, all running at over 1,500 tokens per second.

The practical use cases Cerebras highlights:

• Screenshot Insight: Feed the model a dense dashboard screenshot or document page and get structured output in real time rather than after a wait.
• Long-context summarization: Hand it a research report or technical brief and get a decision-ready summary fast enough to read, react, and re-query in a single sitting.
• Screenshot to Patch: Take a broken UI screenshot, the source code, and the console error , the model returns a minimal patch and verification checks.

Cerebras recommends Gemma 4 as the reference medium-size model on the platform for those building multimodal reasoning, document understanding, fast summarization, or targeted coding workflows where inference speed is the bottleneck.

The Gemma 4 Ecosystem Moment

Thanks to the developer community, Gemma 4 models have now crossed 150 million downloads. Gemma 4 is released under the Apache 2.0 license, which allows free use, modification, and commercial deployment , a change from Gemma 3 and earlier, which used Google's source-available Gemma Terms of Use rather than a fully open-source license. That licensing shift is significant: it means enterprises can fine-tune and deploy without legal ambiguity.

What makes Gemma 4 stand out is its streamlined approach to processing visual and audio inputs. Traditional multimodal models typically rely on separate encoders to translate images and audio before passing those representations to the language model. Because these split encoders add latency and increase memory usage, Google trained Gemma 4 12B with an encoder-free architecture to integrate audio and vision input directly.

The $5,000 Hackathon

To put early access in developers' hands immediately, Cerebras and Google DeepMind are running a 24-hour hackathon. The prize is $5,000, and the winning project gets featured by both Google DeepMind and Cerebras. It is a focused, fast-format competition , the kind designed to generate real prototypes rather than slide decks. RSVP is open via the Cerebras event page on Luma.

What This Means for Your Stack

Cerebras recommends Gemma 4 as the reference medium-size model on the platform: if you are looking to migrate from Llama, GPT-OSS, or Haiku, Gemma 4 provides equal or higher intelligence at Cerebras speed. The combination of Apache 2.0 licensing, 15x speed advantage over comparable hosted models, and now image support makes this a serious option for anyone building agentic pipelines where latency is the bottleneck.

General availability is expected by the end of June. If your application involves any loop where a model needs to see something, reason about it, and act , document processing, UI automation, visual QA, screenshot-driven debugging , this is worth a close look.