Google Gemma 4 Hits the Scene The New Open-Weight Leader in Coding and Multimodality.

Google Unveils Gemma 4: The New Open-Model Powerhouse Dominating Global Benchmarks

Google has officially released Gemma 4, its latest generation of open-weight Large Language Models (LLMs). Designed for high-performance and accessible AI development, the suite features four distinct models: E2B, E4B, 26B-A4B, and 31B. Early evaluations show that Gemma 4 is not just an incremental update, but a massive leap forward in open-source AI capability.

Dominating the Leaderboards

The flagship Gemma 4 31B has made a stunning debut, securing the 27th spot on the Arena.ai (LMSYS) leaderboard. This makes it the 3rd highest-ranked open model globally, trailing only GLM-5 and Kimi K2.5 both of which are significantly larger in scale. Furthermore, the 26B-A4B variant has claimed 6th place in the open-model category, proving that Google’s efficiency optimizations are paying off.

Native Multimodal Power

Every model in the Gemma 4 family features native multimodal support for both images and audio. This integration allows for seamless execution of specialized tasks, such as:

• Audio-to-Text Transcription: High-accuracy speech recognition.

• Visual OCR: Advanced text extraction and understanding from complex images.

The performance gains over Gemma 3 are extraordinary. For instance, in the LiveCodeBench coding benchmark, the score for the 26B-A4B model skyrocketed from 29.1% to 77.1%, while the 31B model reached an elite 80.0%.

Availability and Hardware Support

Gemma 4 is available for immediate download and is optimized to run across a variety of hardware ecosystems, including NVIDIA and AMD GPUs, as well as Google’s own TPU infrastructure.

 

The A4B (Active 4 Billion) designation in the 26B-A4B version indicates the use of a Mixture-of-Experts (MoE) architecture. This means that even though the model has a total of 26 billion parameters, it only uses 4 billion parameters per processing cycle. This allows it to run as "fast and economical" as small-scale models, but as "smart" as large-scale models.

The LiveCodeBench score reaching 80% is phenomenal, as it's comparable to the world's top closed-source models. Analysts believe Gemma 4 will become the new standard for AI coding assistants, allowing developers to run code locally within their own machines without leaking the code outside the company.

Unlike older models that required separate models to handle image and audio, Gemma 4 uses Unified Tokenization. This allows for smooth and very low latency cross-media processing (e.g., watching video and receiving audio responses), making it ideal for real-time AI agents on mobile devices or edge computing.

Google's release of one of the world's top three most powerful open-weighted models directly challenges Meta's Llama 4. This should encourage the developer ecosystem to return to Google's tools after losing market share to Meta in the past year.

 

Anthropic has requested the deletion of leaked data sent via an NPM package. 

 

Source: Google Blog 

💬 AI Content Assistant

Ask me anything about this article. No data is stored for your question.

Google Unveils Gemma 4: The New Open-Model Powerhouse Dominating Global Benchmarks

Google has officially released Gemma 4, its latest generation of open-weight Large Language Models (LLMs). Designed for high-performance and accessible AI development, the suite features four distinct models: E2B, E4B, 26B-A4B, and 31B. Early evaluations show that Gemma 4 is not just an incremental update, but a massive leap forward in open-source AI capability.

Dominating the Leaderboards

The flagship Gemma 4 31B has made a stunning debut, securing the 27th spot on the Arena.ai (LMSYS) leaderboard. This makes it the 3rd highest-ranked open model globally, trailing only GLM-5 and Kimi K2.5 both of which are significantly larger in scale. Furthermore, the 26B-A4B variant has claimed 6th place in the open-model category, proving that Google’s efficiency optimizations are paying off.

Native Multimodal Power

Every model in the Gemma 4 family features native multimodal support for both images and audio. This integration allows for seamless execution of specialized tasks, such as:

• Audio-to-Text Transcription: High-accuracy speech recognition.

• Visual OCR: Advanced text extraction and understanding from complex images.

The performance gains over Gemma 3 are extraordinary. For instance, in the LiveCodeBench coding benchmark, the score for the 26B-A4B model skyrocketed from 29.1% to 77.1%, while the 31B model reached an elite 80.0%.

Availability and Hardware Support

Gemma 4 is available for immediate download and is optimized to run across a variety of hardware ecosystems, including NVIDIA and AMD GPUs, as well as Google’s own TPU infrastructure.

 

The A4B (Active 4 Billion) designation in the 26B-A4B version indicates the use of a Mixture-of-Experts (MoE) architecture. This means that even though the model has a total of 26 billion parameters, it only uses 4 billion parameters per processing cycle. This allows it to run as "fast and economical" as small-scale models, but as "smart" as large-scale models.

The LiveCodeBench score reaching 80% is phenomenal, as it's comparable to the world's top closed-source models. Analysts believe Gemma 4 will become the new standard for AI coding assistants, allowing developers to run code locally within their own machines without leaking the code outside the company.

Unlike older models that required separate models to handle image and audio, Gemma 4 uses Unified Tokenization. This allows for smooth and very low latency cross-media processing (e.g., watching video and receiving audio responses), making it ideal for real-time AI agents on mobile devices or edge computing.

Google's release of one of the world's top three most powerful open-weighted models directly challenges Meta's Llama 4. This should encourage the developer ecosystem to return to Google's tools after losing market share to Meta in the past year.

 

Anthropic has requested the deletion of leaked data sent via an NPM package. 

 

Source: Google Blog