April 3rd news, this morningGoogle DeepMind released a new generationOpen Source ModelSeries Gemma 4, one-time roll-out of four models covering the end-to-end to the workstation scene。
- E2B: 5.1 BILLION TOTAL PARAMETERS, 2.3 BILLION VALID PARAMETERS, 128K CONTEXT, OFFICIALLY STATED THAT SOME EQUIPMENT MEMORY OCCUPANCY CAN BE PRESSED BELOW 1.5GB
- E4B: 8 billion total parameters, 4.5 billion valid parameters, 128K context, MMLLU Pro 69.4%, close to previous generation 27B level
- 26B A4B MoE: 25.2 billion total parameters, only 3.8 billion activated parameters, reasoning speed close to 4B model, Arena AI Open SourceNumber six
- 31B Dense: 31 Billion Parameters All Activated, 256K context, and Arena AI third in the open list。
The performance aspect is significantly higher than the previous Gemma 3 27B. 31B Scored 89.2% (update 208%), LiveCodeBench v6 code test jumped from 29.1% to 80.0%, long document processing MRRCR v2 128K from 13.5% to 66.4%。
All models support the input of images and videos and 140 multiple languages, and include the thought mode of switches. E2B and E4B additional audio encoders with approximately 300 million parameters, which support voice recognition and translation for up to 30 seconds, can be fully offline on mobile phones, berry pies and NVIDIA Jetson Orin Nano。
It is worth noting that the Gemma series open source agreement is now fully converted to Apache 2.0, with developers free to modify, distribute and commercialize without user thresholds. The co-founder of Hugging Face, Clément Delange, is evaluated as a "significant milestone"。