Google Launches Gemma4 Open-Source Model in Four Specs, 31B Version Ranks Third Globally
At 4:03 AM Beijing Time on April 3rd, Google officially launched the open-source large language model Gemma4. Featuring a breakthrough in "parameter efficiency per unit," it sets a new benchmark for open-source models in powering intelligent agent workflows.
The series includes the efficient E2B (2.3B) and E4B (4.5B) variants, alongside the high-performance 26B MoE and 31B dense models. As the latest iteration built on the Gemini3 technology stack, Gemma4 offers full multimodal input support (images and videos). The E2B and E4B models also natively support voice input, enabling real-time voice understanding at the edge.

In terms of technical architecture, the large-parameter models achieve exceptional hardware efficiency through optimization. The 31B dense version ranks third globally among open-source models on the Arena AI text leaderboard, while the 26B MoE version holds the sixth position. Its logical reasoning and function calling capabilities are robust enough to power complex autonomous agents.
For local deployment, Gemma4 dramatically lowers the barrier to accessing cutting-edge AI. The non-quantized 31B model weights can run on a single 80GB H100 GPU, while quantized versions are compatible with consumer-grade GPUs. For mobile and IoT devices, the E2B and E4B models achieve low-latency logical processing on platforms like Raspberry Pi and smartphones, thanks to innovative PLE embedding technology and 128K context length support.
This release not only showcases Google's deep commitment to the open-source ecosystem but also, through its Apache 2.0 licensing, provides a foundation for developers worldwide to build localized, high-privacy AI applications.
Related article
Anthropic Expands Claude AI Coding Tools to Japan in Push for Overseas Growth
Anthropic, a prominent U.S. artificial intelligence firm, is intensifying its global outreach. On Wednesday, the company hosted a major developer gathering, "Code with Claude," in Tokyo, drawing close to 500 software engineers. This initiative seeks
India’s Software Giant Cuts Hiring, Promises No Layoffs as AI Agents Scale Up
As artificial intelligence reshapes traditional labor-intensive business models, Tata Consultancy Services (TCS), a premier Indian software outsourcing firm, has unveiled its strategic response. During Tuesday’s annual general meeting, the chairman o
China Locks AI Models During Gaokam to Block Instant Homework Help
With the 2026 Gaokao fast approaching, rumors regarding the suspension of AI tools during the exam period have ignited intense online debate. In response to public concern, major AI platforms and educational apps have clarified their stance: rather t
Related Special Topic Recommendations
Comments (2)
0/500
Finally Google open-sources something decent! The 31B version ranking third globally is impressive, but I wonder how it compares to Llama 3 in real-world agent tasks? Honestly, parameter efficiency sounds like marketing fluff, but let's see the benchmarks. 🤔
At 4:03 AM Beijing Time on April 3rd, Google officially launched the open-source large language model Gemma4. Featuring a breakthrough in "parameter efficiency per unit," it sets a new benchmark for open-source models in powering intelligent agent workflows.
The series includes the efficient E2B (2.3B) and E4B (4.5B) variants, alongside the high-performance 26B MoE and 31B dense models. As the latest iteration built on the Gemini3 technology stack, Gemma4 offers full multimodal input support (images and videos). The E2B and E4B models also natively support voice input, enabling real-time voice understanding at the edge.

In terms of technical architecture, the large-parameter models achieve exceptional hardware efficiency through optimization. The 31B dense version ranks third globally among open-source models on the Arena AI text leaderboard, while the 26B MoE version holds the sixth position. Its logical reasoning and function calling capabilities are robust enough to power complex autonomous agents.
For local deployment, Gemma4 dramatically lowers the barrier to accessing cutting-edge AI. The non-quantized 31B model weights can run on a single 80GB H100 GPU, while quantized versions are compatible with consumer-grade GPUs. For mobile and IoT devices, the E2B and E4B models achieve low-latency logical processing on platforms like Raspberry Pi and smartphones, thanks to innovative PLE embedding technology and 128K context length support.
This release not only showcases Google's deep commitment to the open-source ecosystem but also, through its Apache 2.0 licensing, provides a foundation for developers worldwide to build localized, high-privacy AI applications.
Anthropic Expands Claude AI Coding Tools to Japan in Push for Overseas Growth
Anthropic, a prominent U.S. artificial intelligence firm, is intensifying its global outreach. On Wednesday, the company hosted a major developer gathering, "Code with Claude," in Tokyo, drawing close to 500 software engineers. This initiative seeks
India’s Software Giant Cuts Hiring, Promises No Layoffs as AI Agents Scale Up
As artificial intelligence reshapes traditional labor-intensive business models, Tata Consultancy Services (TCS), a premier Indian software outsourcing firm, has unveiled its strategic response. During Tuesday’s annual general meeting, the chairman o
China Locks AI Models During Gaokam to Block Instant Homework Help
With the 2026 Gaokao fast approaching, rumors regarding the suspension of AI tools during the exam period have ignited intense online debate. In response to public concern, major AI platforms and educational apps have clarified their stance: rather t
Finally Google open-sources something decent! The 31B version ranking third globally is impressive, but I wonder how it compares to Llama 3 in real-world agent tasks? Honestly, parameter efficiency sounds like marketing fluff, but let's see the benchmarks. 🤔





Home






