Google Lanca Gemma 4 Open-Source and Gemini 3.1 Ultra Break Records

Google did in April 2026 what many analysts did not expect: it launched Gemma 4 completely open-source in four different sizes and, in the same announcement, revealed that Gemini 3.1 Ultra achieved 94.3% in GPQA Diamond -- the industry's most difficult benchmark for scientific reasoning. It is not a marketing advertisement. And a real shift in the balance of power in AI.

This article analyzes each new feature, explains what it means in practice for those who develop software and those who work in digital marketing, and connects everything with the tools you already use on a daily basis.

1. The big picture: Google attacks AI again

In the last 12 months, the market narrative was clear: Anthropic led in code agents (Claude Code), OpenAI dominated in user base (ChatGPT/Codex) and Meta advanced in open-source (Llama). Google seemed to lag behind, with Gemini being good but not exceptional in any category.

The April package changes this perception. Google attacked on two simultaneous fronts:

• Open source:Gemma 4 is, technically, the best open-source model in the world in the 2B to 31B tometer range. This directly challenges the Meta Llama
• Closed-source premium:Gemini 3.1 Ultra is now the best-scoring model in scientific reasoning benchmarks, surpassing Claude Opus and GPT-5.4
• Practical application:Integrations into Google Marketing Platform, Gmail, Drive and voice assistant show that Google wants AI in every product, not just the API

The strategy is clear: dominate the open-source market with Gemma (capturing developers and startups) while competing at the top with Gemini (capturing companies and premium users). Let's analyze each piece.

2. Gemma 4: open-source in 4 sizes

Gemma 4 is the fourth generation of Google's family of open-source models. The big news is that there are now four variants, each optimized for a different scenario:

The 4 sizes

Model	Parameters	Architecture	Optimal use
Gemma 4 E2B	2 billion	Dense	Smartphones, IoT, edge devices
Gemma 4 E4B	4 billion	Dense	Laptops, desktop applications, lightweight chatbots
Gemma 4 26B MoE	26 billion (MoE)	Mixture of Experts	Servers, APIs, complex tasks efficiently
Gemma 4 31B Dense	31 billion	Dense	Maximum performance, search, advanced fine-tuning

What is Mixture of Experts (MoE)

The 26B MoE model deserves explanation. In MoE architectures, the model has 26 billion tometers in total, but only a fraction of them are activated for each token processed. Think of it like this: instead of one expert who knows everything, you have a team of experts and, for each task, only the relevant ones are called.

In practice, the 26B MoE has comtoble performance to the 31B Dense in most tasks, but uses less memory and processes faster because it does not activate all tometers at the same time. It is the ideal choice for those who want to deploy in production with a good cost-benefit ratio.

Licensing

All four models are distributed under Google's open license, which allows:

• Royalty-free commercial use
• Fine-tuning with your own data
• Redistribution (you can include it in your products)
• Architecture modification

The only relevant restriction is that you cannot use the templates to generate content that violates Google's usage policies (misinformation, illegal content, etc.). For 99% of enterprise and development use cases, this is not a limitation.

3. Gemma 4 benchmarks and what they mean in practice

Benchmarks are useful when contextualized. Here's how the Gemma 4 compares to similarly sized models:

Gemma 4 31B Dense vs competitors

Benchmark	Gemma 4 31B	Llama 3.3 33B	Qwen 3 32B
MMLU (General Knowledge)	84.7%	82.1%	83.2%
HumanEval (code)	81.3%	76.8%	79.1%
GSM8K (math)	92.1%	88.4%	90.6%
GPQA (scientific reasoning)	58.2%	51.7%	54.3%
MT-Bench (conversational)	8.9/10	8.4/10	8.7/10

The numbers show that the Gemma 4 31B is the best open-source model in the 30B tometer range in practically all categories. The advantage is not overwhelming, but it is consistent -- 2 to 6 percentage points above the Llama 3.3 in each benchmark.

Gemma 4 E2B: what impresses

The really surprising model is E2B (2 billion tometers). In code and reasoning benchmarks, it matches or surpasses previous generation 7B tometer models. This means that a model that runs on an Android smartphone achieves performance that, 18 months ago, required a server with a GPU.

For mobile developers, this opens up real possibilities: offline code wizards, intelligent autocomplete without a connection to the cloud, and natural language processing in apps that work without the internet.

4. Gemma 4 on Android and edge devices

Google didn't launch Gemma 4 E2B just as an academic curiosity. There is a direct integration with the Android ecosystem that deserves attention.

Android AI Core

Android AI Core is Google's framework for running AI models locally on Android devices. With Gemma 4 E2B, any Android app can:

• Process text locally:summarize emails, generate responses, translate -- all without sending data to the cloud
• Smart autocomplete:text suggestions in any input field, based on the context of the conversation
• Image analysis:o E2B and multimodal -- understands text and images. Can describe photos, read documents and extract information from screenshots
• Offline Assistant:works without internet, with millisecond latency

Hardware Requirements

The Gemma 4 E2B runs on any smartphone released from 2024 onwards with at least 4GB of RAM. The model occupies around 1.5GB of storage in quantized format (INT4). On a Pixel 8 or Galaxy S24, inference takes less than 200ms per short response.

For comparison: Gemma 3 E2B required almost twice as much memory and was 40% slower. Optimizing Gemma 4 for mobile hardware is real, not just marketing.

Implications for app developers

The race is now on to integrate local AI into existing apps. Smart keyboards, email apps, productivity tools, health apps, education -- any app that handles text or images can benefit from a 2B tometer model running locally. The inference cost is zero (it runs on the user's device) and privacy is total (data never leaves the cell phone).

5. Gemini 3.1 Ultra: 94.3% on GPQA Diamond

If Gemma 4 is the open-source game, Gemini 3.1 Ultra is the premium game. And the numbers are impressive.

What is GPQA Diamond

GPQA Diamond is a scientific reasoning benchmark considered the most difficult in the industry. Questions are created by PhDs and require multi-step reasoning in physics, chemistry, biology and advanced mathematics. For context: human experts (with a PhD in the field) get around 81% of the questions right. Non-expert humans get about 34% right.

Gemini 3.1 Ultra achieved 94.3%. This not only outperforms any other AI model -- it outperforms the average human expert by more than 13 percentage points.

Comparison with the competition

Model	GPQA Diamond	MMLU-Pro	HumanEval
Gemini 3.1 Ultra	94.3%	91.8%	93.2%
Claude Opus (April 2026)	89.7%	90.2%	94.1%
GPT-5.4	87.2%	89.5%	91.8%
Gemini 3.0 Ultra	82.1%	86.4%	88.7%

Gemini 3.1 Ultra leads in GPQA Diamond and MMLU-Pro (advanced general knowledge). Claude Opus continues to lead in HumanEval (code generation), which makes sense -- Anthropic optimizes Opus specifically for coding tasks.

What 94.3% on GPQA Diamond means in practice

For most users, this benchmark does not change their daily lives. You won't feel any difference when asking Gemini to write an email or summarize a document. The difference appears in tasks that require deep reasoning:

• Complex data analysis:the model can identify patterns and correlations in datasets that other models miss
• Scientific code debugging:For researchers working with scientific Python (numpy, scipy, pandas), Gemini 3.1 Ultra understands the mathematical context behind the code
• Strategic planning:the ability to reason in multiple steps translates into better quality of analysis for complex business decisions

Claude Code remains leader for code agents

None of Google's new features replace Claude Code for agent-assisted development. Gemma 4 and Gemini 3.1 are models -- not agents. They don't read your files, they don't execute commands, they don't create projects. To do this, you still need a tool like Claude Code (or Codex) that orchestrates the model as an agent.

The connection between the two worlds: you can use Claude Code with specialized skills for your workflow, and use Gemma 4 or Flash-Lite for specific processing tasks that don't need a full agent.

11. Impact for marketers

If you work in digital marketing, Google's April package brings direct changes to your daily life:

Google Marketing Platform with Gemini

The most awaited integration. If you manage campaigns on Google Ads, the copy generation and automatic diagnosis tools will save you hours per week. The key is to use Gemini as an accelerator, not a replacement -- review everything before publishing.

Personal Intelligence for productivity

If you use Gmail and Google Drive (and who doesn't?), the semantic search and document summary features are immediately useful. Instead of opening 5 spreadsheets to create a monthly report, ask Gemini and it will consolidate.

Flash Live for costmer service

If you manage costmer service or support, Flash Live can be integrated as the first level of voice support. The 300ms latency and multimodal capability (the costmer can show a product through the camera) create an experience that previously required human agents.

Skills + Gemini: the ideal combination

For those who already use Claude Code with marketing skills, Gemini on Google Marketing Platform complements, not replaces. Use Claude Code with skills to create landing pages, configure tracking and generate long copy. Use Gemini in GMP for campaign optimization, diagnostics and short ad creatives.

The trend is clear: marketers who master multiple AI tools (not just one) will have a competitive advantage. It's not about choosing Gemini or Claude Code -- it's about using each where they're best.