Trending Now - Google Introduces Gemma 4: Advanced Open Models for Reasoning

WHO:

Techmeme - Google, specifically the Group Product Manager from Google DeepMind, announced the launch of Gemma 4.

Engadget - Google is the organization releasing the Gemma 4 models.

WHAT:

Techmeme - Google claims that Gemma 4 is "the most intelligent open model family" it has released to date.

Engadget - The 31-billion dense model of Gemma 4 achieved a score of 89.2% on AIME 2026, 80.0% on LiveCodeBench v6, and a Codeforces ELO of 2,150.

VentureBeat - The 26-billion A4B Mixture-of-Experts model scored 88.3% on AIME 2026, 77.1% on LiveCodeBench, and 82.3% on GPQA Diamond.

VentureBeat - The E4B model scored 42.5% on AIME 2026 and 52.0% on LiveCodeBench, while the E2B model scored 37.5% and 44.0% respectively.

VentureBeat - The 26-billion A4B model activates only 3.8 billion of its 25.2 billion total parameters during inference.

VentureBeat - The audio encoder in the edge models has been compressed to 305 million parameters, down from 681 million in Gemma 3.

VentureBeat - The frame duration for audio processing dropped from 160ms to 40ms for more responsive transcription.

VentureBeat - Google is releasing both pre-trained base models and instruction-tuned variants for the Gemma 4 family.

Ars Technica - Gemma 4 brings the first major update to Google's open models in a year.

Techmeme - Google stated that Gemma 4 is purpose-built for advanced reasoning and agentic workflows.

VentureBeat - The 31-billion dense model of Gemma 4 scores 89.2% on AIME 2026, 80.0% on LiveCodeBench v6, and achieves a Codeforces ELO of 2,150.

VentureBeat - The 26-billion A4B Mixture-of-Experts model scores 88.3% on AIME 2026, 77.1% on LiveCodeBench, and 82.3% on GPQA Diamond.

VentureBeat - The E4B model scores 42.5% on AIME 2026 and 52.0% on LiveCodeBench, while the E2B model scores 37.5% and 44.0% respectively.

VentureBeat - The audio encoder in the edge models has been compressed to 305 million parameters, down from 681 million in Gemma 3.

VentureBeat - The frame duration for audio processing dropped from 160ms to 40ms for more responsive transcription.

Techmeme - Google claims that the Gemma 4 family represents "an unprecedented level of intelligence-per-parameter."

VentureBeat - The 31-billion dense model of Gemma 4 achieved a Codeforces ELO of 2,150, which is a significant performance metric.

VentureBeat - The 26-billion A4B Mixture-of-Experts model activates only 3.8 billion of its 25.2 billion total parameters during inference.

VentureBeat - The E4B model scores 42.5% on AIME 2026 and 52.0% on LiveCodeBench, while the E2B model scores 37.5% and 44.0% respectively.

VentureBeat - The 31-billion dense model requires more powerful hardware, such as an NVIDIA H100 or RTX 6000 Pro for unquantized inference.

VentureBeat - Google is shipping Quantization-Aware Training (QAT) checkpoints to maintain quality at lower precision for the workstation models.

VentureBeat - The 26-billion A4B model is designed to run on consumer-grade GPUs, while the 31-billion dense model requires more headroom.

VentureBeat - The MoE model can run on consumer-grade GPUs and should appear quickly in tools like Ollama and LM Studio.

VentureBeat - The two edge models add native audio processing, including automatic speech recognition and speech-to-translated-text, all on-device.

VentureBeat - The audio encoder in the edge models has been compressed to 305 million parameters, down from 681 million in Gemma 3.

VentureBeat - The frame duration for audio processing dropped from 160ms to 40ms for more responsive transcription.

VentureBeat - Google is releasing both pre-trained base models and instruction-tuned variants for the Gemma 4 family.

Ars Technica - Gemma 4 brings the first major update to Google's open models in a year.

VentureBeat - The 31-billion dense model scores 89.2% on AIME 2026, 80.0% on LiveCodeBench v6, and hits a Codeforces ELO of 2,150.

VentureBeat - The 26-billion A4B Mixture-of-Experts model scores 88.3% on AIME 2026, 77.1% on LiveCodeBench, and 82.3% on GPQA Diamond.

VentureBeat - The E4B model scores 42.5% on AIME 2026 and 52.0% on LiveCodeBench, while the E2B model scores 37.5% and 44.0% respectively.

Techmeme - Google described Gemma 4 as purpose-built for advanced reasoning and agentic workflows.

Engadget - The 2-billion and 4-billion models of Gemma 4 are referred to as "Effective" models, while the 26-billion and 31-billion models are categorized as "Mixture of Experts" and "Dense" systems, respectively.

Engadget - Google has trained the Gemma 4 models in more than 140 languages.

Techmeme - Google launched Gemma 4, described as its "most intelligent" open model family.

Engadget - The Gemma 4 family includes four different versions differentiated by the number of parameters: 2-billion, 4-billion, 26-billion, and 31-billion.

Engadget - The 31-billion and 26-billion variants of Gemma 4 ranked third and sixth, respectively, on Arena AI's text leaderboard, outperforming models 20 times their size.

Engadget - All models in the Gemma 4 family can process video and images, and the two smaller models can process audio inputs and understand speech.

Engadget - Gemma 4 models are capable of generating offline code.

WHERE:

Engadget - The model weights for Gemma 4 are available through platforms such as Hugging Face, Kaggle, and Ollama.

WHEN:

Engadget - Google released Gemini 3 Pro at the end of 2025, prior to the launch of Gemma 4 in 2026.

WHY:

VentureBeat - The switch to an Apache 2.0 license eliminates legal friction for enterprises evaluating open-weight models, allowing for greater freedom in modification and deployment.

Engadget - Google stated that the open-source license (Apache 2.0) allows for complete control over data, infrastructure, and models, promoting developer flexibility and digital sovereignty.

Engadget - Google aims to provide the open-source community with advanced reasoning and agentic workflows through the release of Gemma 4.

Engadget - The open-source license (Apache 2.0) allows for greater freedom to modify the models, promoting developer flexibility and digital sovereignty.

HOW:

VentureBeat - The architectural choices inside the 26-billion A4B model utilize 128 small experts, activating eight per token plus one shared always-on expert, which affects serving costs and inference economics.

VentureBeat - Both workstation models use a hybrid attention mechanism that interleaves local sliding window attention with full global attention, with the final layer always global.

VentureBeat - The 26-billion A4B model activates only 3.8 billion of its 25.2 billion total parameters during inference.

VentureBeat - The 31-billion dense model requires more powerful hardware, such as an NVIDIA H100 or RTX 6000 Pro for unquantized inference.

VentureBeat - Google is shipping Quantization-Aware Training (QAT) checkpoints to maintain quality at lower precision for the workstation models.

Engadget - Google claims to have achieved "an unprecedented level of intelligence-per-parameter" with the Gemma 4 models.

Engadget - Google engineered the Gemma 4 models to achieve "an unprecedented level of intelligence-per-parameter."

(UPDATED April 3, 2026 03:51 UTC)

Google has announced that its new Gemma 4 model family is the most intelligent open model it has released to date, achieving impressive scores such as 89.2% on AIME 2026 and a Codeforces ELO of 2,150 for its 31-billion dense model. The Gemma 4 family includes pre-trained base models and instruction-tuned variants, with model weights available on platforms like Hugging Face and Kaggle, and the switch to an Apache 2.0 license aims to reduce legal barriers for enterprises. Additionally, the architectural design of the 26-billion A4B model features 128 small experts that optimize inference costs, while the audio encoder has been significantly compressed for improved performance.

(UPDATED April 3, 2026 01:21 UTC)

Google has announced that its new Gemma 4 model is specifically designed for advanced reasoning and agentic workflows, achieving impressive scores on various benchmarks. The 31-billion dense model scores 89.2% on AIME 2026 and has a Codeforces ELO of 2,150, while the 26-billion A4B model also performs well with scores of 88.3% on AIME 2026 and 77.1% on LiveCodeBench. Additionally, the switch to an Apache 2.0 license is expected to reduce legal barriers for enterprises, promoting greater flexibility in the use of open-weight models.

(UPDATED April 2, 2026 23:09 UTC)

Google has announced that its new Gemma 4 family boasts what they describe as "an unprecedented level of intelligence-per-parameter," with the 31-billion dense model achieving a Codeforces ELO of 2,150. The 26-billion A4B Mixture-of-Experts model is designed for consumer-grade GPUs and activates only a fraction of its parameters during inference, while both models are set to include advanced audio processing features. Additionally, the switch to an Apache 2.0 license aims to reduce legal barriers for enterprises looking to utilize these open-weight models, enhancing their ability to modify and deploy the technology.

(UPDATED April 2, 2026 21:08 UTC)

Google has released the first major update to its open models in a year, introducing the Gemma 4. The new 31-billion dense model has achieved impressive scores of 89.2% on AIME 2026 and 80.0% on LiveCodeBench v6, while the 26-billion A4B Mixture-of-Experts model scored 88.3% on AIME 2026 and 77.1% on LiveCodeBench. Additionally, Google is implementing Quantization-Aware Training checkpoints to ensure quality at lower precision for its workstation models, with the 26-billion A4B model activating only a fraction of its total parameters during inference.

(UPDATED April 2, 2026 19:20 UTC)

Google has unveiled its latest AI advancement, Gemma 4, which is designed for advanced reasoning and agentic workflows, featuring models that range from 2 billion to 31 billion parameters. The company has trained these models in over 140 languages and categorizes them into "Effective," "Mixture of Experts," and "Dense" systems, depending on their size and capabilities. Google emphasizes that the open-source license allows developers complete control over their data and infrastructure, promoting flexibility and digital sovereignty.

(UPDATED April 2, 2026 17:20 UTC)

Google has officially launched Gemma 4, which is being touted as its most intelligent open model family, featuring four versions with parameters ranging from 2 billion to 31 billion. The 31-billion and 26-billion variants have already made a mark by ranking third and sixth on Arena AI's text leaderboard, outperforming models that are 20 times their size, and all models can process video and images, with the smaller ones also capable of handling audio inputs. Available through platforms like Hugging Face and Kaggle, the Gemma 4 models are designed to enhance the open-source community with advanced reasoning capabilities, all under an open-source license that encourages developer flexibility.