Google announced Wednesday JEMA 3The next repetition in the Jima family of openweight models. It is the successor of the JEMA 2 model released last year.
The smaller model comes to a range of parameter -shaped range – 1b, 4b, 12b and 27b. The model also supports the long context of 128K tokens. It can analyze videos, photos and text, supports 35 languages outside the box, and provides pre -trained support for 140 languages.
In the Chat Boat Arena, Jima 37B improved the Deep Sak-V 3, Openi’s O3-Mini and Meta Lama 3-405B model. In the Chat Boat Arena, models are evaluated against each other against each other through diagnosis.
Further, Jema 3 27B scored 67.5 % and 42.4 Standard benchmark Such as MMLU-PRO, GPQA Diamond. The model performs well than other smaller models of competition.
Claude 3.5 Haiku scored 63 % on MMLU Pro Benchmark and 41 % on GPQA Diamond, while Openi GPT -4OO Money got 65 % and 43 % respectively. Meta Lama 3.3 70B performed both well, with 71 % in MMLU Pro and 50 % in GPQA Diamond, making it the strongest contender in these models.
However, it seems that JEMA 3’s key superpower is efficient use. Google said Jema 327B scored a score with a single Nvidia H100 GPU, while other models need up to 32 GPUs.
Source: Google
The company also revealed that the architecture of the model was changed to reduce KV catch memory, which increases the long context.
Google has a Published a detailed technical report Exercise the techniques used to build the model, its performance and other features. Jema 3 can be accessed through different ways. Google is offering a model on the web using it Google AI StudioBy default chat boot or by API, and it is also available Google Geni SD’s.
In addition, the model can be downloaded for local deployment The hugs faceFor, for, for,. BesetAnd Cogl.
Along with JEMA 3, Google has also launched Shield Jama 2, which is built on the Foundation of 4B Parameter Image Safety Checker Jima 3. It provides a protective label for harmful images, including dangerous, sexually clear and violent material.
