First came DeepSeek, from China, close to the Chinese New Year, in January 2025. It was a fact that DeepSeek has been developed by spending just $5.6 million, and that came as a slap across the face for the so-called ‘Mughuls’ of the global tech industry. Its R1 is able to match the most powerful AI models, such as those from ChatGPT, Meta, and Google!
Well, guess what! Google has now come up with Gemma 3! Interestingly. Gemma 3 27B ranks quite highly on a chart that ranks AI models. Gemma 3 is said to be requiring only a single GPU, despite others needing up to 32! That’s another engineering marvel! By the way, Google did not release any statement regarding how much they spent on developing the Gemma 3. If they did, sorry, I missed that.
As per the chart mentioned above, it seems that R1 is still superior to Gemma 3, albeit, by a very narrow margin — In the chatbot Arena Elo score, R1 has score of 1363, while Gemma 3 has a score of 1338.
As per Google, Gemma 3 introduces multimodality, supporting vision-language input and text outputs. It handles context windows up to 128k tokens, understands over 140 languages, and offers improved math, reasoning, and chat capabilities, including structured outputs and function calling. Gemma 3 is available in four sizes (1B, 4B, 12B, and 27B) as both pre-trained models, which can be fine-tuned for your own use cases and domains, and general-purpose instruction-tuned versions.
Any idea how many languages does DeepSeek support? Also, the dividing line can be how DeepSeek and Gemma 3 handle language responses, other than English and Chinese, and how well they can provide information. Well, these are all works in progress, with more to come.
Gemma’s pre-training and post-training processes were optimized using a combination of distillation, reinforcement learning, and model merging. This approach results in enhanced performance in math, coding, and instruction following. Gemma 3 uses a new tokenizer for better multilingual support for over 140+ languages and was trained on 2T tokens for 1B, 4T for 4B, 12T for 12B, and 14T tokens for 27B, on Google TPUs using the JAX Framework.
For post-training, Gemma 3 uses 4 components:
* Distillation from a larger instruct model into the Gemma 3 pre-trained checkpoints.
* Reinforcement Learning from Human Feedback (RLHF) to align model predictions with human preferences.
* Reinforcement Learning from Machine Feedback (RLMF) to enhance mathematical reasoning.
* Reinforcement Learning from Execution Feedback (RLEF) to improve coding capabilities.
These updates significantly improved the model math, coding, and instruction following capabilities, making it the top open compact model in LMArena, with a score of 1338.
With a fully on-device Gemma 3 1B model, you are able to take advantage of the benefits of AI edge. These are:
Offline availability: Enable your app to work fully when WiFi or cellular data is unavailable.
Cost: With no cloud bills, enable free or freemium apps.
Latency: Some features need to be faster than a server call allows.
Privacy: Bring intelligence to data that is unable to leave the device or is end-to-end encrypted.
Gemma 3 is available for download through platforms such as Kaggle and Hugging Face. It is also accessible via Google Studio.
Once can only appreciate the great work done by DeepSeek and Google in developing their respective AI models. And, where is OpenAI? Well, right now, nowhere! Can it come back? Surely, yes!
Talking on CNBC, Deirdre Bosa, said Gemma 3 can be run on a single Nvidia H100 vs. the equivalent of 34 of them needed to run deep. This is said to be more accessible and cost effective for enterprises. Google also has a chip efficiency edge here. Hyperscalers are also said to be developing their own in-house AI chips to reduce costs and remove reliance on Nvidia.
Efficiency is now shifting the AI trade. There is the growing Nvidia threats. Anthropic is using Google TPUs. Amazon Trainium chips are making inroads. Meta has begun testing its own in-house chip. OpenAI has very big ambitions here. Perplexity is using Cerebras chips.
Startups are also gaining traction. Cerebras has specialized AI chips. They are powering Perplexity, SONAR model, and Mistral’s chatbot. All of this is raising questions over Nvidia’s dominance in the next phase. If other options are finding more efficiency, and are able to offer lower costs, that can be game changing. Economics of AI are surely said to be shifting. There is lot more competition.
No matter, someone else, or something new, will overtake these two, and probably, very soon. Come on guys, we are already in the AI race! And, it is getting hotter, really well!
Regarding India, there was a Summit recently, where there was lot of talk about AI. Well, good to see you all keep talking! Now, better start working too! There is much more happening elsewhere.
