Notes on Gemma 3
Some quick hands-on check up for the new Gemma 3 model, with 1 billion parameter locally with olllama
Google announced a new family of models, which are open-sourced weights and have 4 variations of the parameters, 1, 4, 12, and 27 billion parameters. Leaving the 1 billion parameter, the rest 3 parameters are multi-modal, they support images and short videos.
The model card for gemma 3: https://ai.google.dev/gemma/docs/core/model_card_3
The announcement blog: https://blog.google/technology/developers/gemma-3/
The developer guide: https://developers.googleblog.com/en/introducing-gemma3/
The below notes are for Gemma 3 on a 1 billion model running locally via Ollama.
Quick snippet to try out gemma 3:1b on Ollama
curl -fsSL https://ollama.com/install.sh | sh
ollama pull gemma3:1b
ollam run gemma3:1b
Tool Calling
The models seem to have a functional call, which could open up the possibility of tool calling via json mode. This is a thing that is driving the current wave of agentic workflows and LLMs. So, if any LLM which is new and quite lightweight, having tool calling ability makes it a superpower. This makes an LLM more powerful in adoption and integration into new or existing workflows I think.
Context Length
The smallest model (1b) has a 32k context window, which is quite impressive. This is compared to Llama 3.2, which had only 8k of context window. So, that’s a 4x jump in the input tokens.
Also to note, the rest of the versions 4B, 12B, and 27B have a context window of 128k, which is the standard right now set by OpenAI, and guess, everyone is trying to keep up with it.
Pre-Trained
The model seems to have a knowledge cutoff around November 2024. Since its a pre-trained model, that is expected.
Simple tests
It did okayish, not bad, but not too good either.
I mean, I know it's a one-billion-parameter model, but that’s still an LLM predicting the next token. It will say there are 4 or 2 Rs in stRawbeRRy.
I did go the extra mile and let it have a moment of life moment
>>> how many Rs in stRawbeRRy?
The value of `Rs` in `stRawbeRRy` is **0**.
It’s a placeholder value, often used to represent a missing or undefined value. It’s a common practice in programming to use a placeholder when a value isn’t available.
>>> how many 'R's in stRawbeRRy?
There are **two** ‘R’s in `stRawbeRRy`.
>>> really?
You are absolutely right to question me! My apologies. I made a mistake.
There are **three** ‘R’s in `stRawbeRRy`. I was processing the input incorrectly. Thanks for catching that!
Let me know if you have any other questions.
>>> really?
You are absolutely right to call me out! I am still under development and sometimes make mistakes. I apologize for the repeated error.
There are **three** ‘R’s in `stRawbeRRy`.
I appreciate you pointing that out – it helps me learn!
>>> are you sure?
You are absolutely right to question me! You are correct. I am *certain* there are only three ‘R’s in `stRawbeRRy`. I am still under development and learning to avoid these errors. Thank you for your patience!
>>> what if 1 == 2?
You are absolutely correct! My apologies again – I’m still struggling with this.
If 1 == 2, then `stRawbeRRy` would be equal to 4.
Thanks for keeping me honest!
That is a quick way of cracking jokes. Its not able to understand and replies in a hurried and not-so logical way. (Google seems to be having data cleaning issues).
This did a little better, but not up to the mark.
Llama3.2 1 billion is much better than that, at least it’s close
Citations
The model tries its best to hallucinate while doing citations, its like those things where it just is making up links from the words you threw at the input.
Most of the links were broken, especially the detailed ones, right? The YouTube channels, the Slack channel, and the specific courses. The documentation links are quite easy to make up; however, once you drill down on a thing, it sort of messes up words and gambles.
>> list some great resources on learning kubernetes
Okay, let's dive into some fantastic resources for learning Kubernetes! There's a lot of
information out there, so I've broken it down into categories to help you find what best suits
your learning style and experience level.
**1. Official Kubernetes Documentation (The Foundation)**
* **Kubernetes.io:** ([https://kubernetes.io/](https://kubernetes.io/)) – This is *the*
official source. It’s comprehensive, well-maintained, and constantly updated. It covers
everything from basic concepts to advanced topics.
* **Concepts:**
[https://kubernetes.io/docs/concepts/](https://kubernetes.io/docs/concepts/) - A great
starting point for understanding the core concepts.
* **Getting Started:**
[https://kubernetes.io/docs/getting-started/](https://kubernetes.io/docs/getting-started/) -
A practical guide to setting up your first Kubernetes cluster.
* **Kubernetes Tutorials:**
[https://kubernetes.io/docs/tutorials/](https://kubernetes.io/docs/tutorials/) - A collection
of tutorials covering various topics.
* **Kubernetes Documentation - Concepts:**
[https://kubernetes.io/docs/concepts/](https://kubernetes.io/docs/concepts/) - A detailed
breakdown of Kubernetes concepts.
**2. Interactive Tutorials & Courses (Hands-on Learning)**
* **Katacoda:**
([https://katacoda.com/courses/kubernetes](https://katacoda.com/courses/kubernetes)) –
Katacoda offers interactive, browser-based courses that let you build Kubernetes clusters and
applications directly. Excellent for beginners and those who prefer a hands-on approach.
They have free and paid options.
* **A Cloud Guru Kubernetes Course:**
([https://www.cloudguru.com/kubernetes-course/](https://www.cloudguru.com/kubernetes-course/))
– A popular paid course that covers the fundamentals and more advanced topics. It's
well-structured and includes labs.
* **Udemy - Kubernetes for Beginners:**
([https://www.udemy.com/course/kubernetes-for-beginners/](https://www.udemy.com/course/kubernet([https://www.udemy.com/course/kubernetes-for-beginners](https://www.udemy.com/course/kubernetes-for-beginners/)) – A more affordable option with a good range of reviews and a focus on
practical application.
* **Linux Foundation Training - Kubernetes:**
([https://training.linuxfoundation/](https://training.linuxfoundation/)) – Offers a structured
learning path with certifications. Good for those looking for a more formal education.
* **Codecademy - Learn Kubernetes:**
([https://www.codecademy.com/learn/learn-kubernetes](https://www.codecademy.com/learn/learn-kub([https://www.codecademy.com/learn/learn-kubernetes](https//www.codecademy.com/learn/learn-kubernetes)) – A more beginner-friendly interactive course that introduces you to the basics.
**3. YouTube Channels**
* **Kubernetes.io YouTube Channel:**
([https://www.youtube.com/@KubernetesIO](https://www.youtube.com/@KubernetesIO)) - Official
Kubernetes channel with tutorials, announcements, and community content.
* **System76:** ([https://www.youtube.com/@System76](https://www.youtube.com/@System76)) -
Excellent for learning Kubernetes through practical examples and troubleshooting. They have a
lot of focused tutorials.
* **FreeCodeCamp:**
([https://www.youtube.com/@FreeCodeCamp](https://www.youtube.com/@FreeCodeCamp)) - They have
a Kubernetes course that is free and covers a wide range of topics.
**4. Books**
* **"Kubernetes in Action" by Michael Scott:**
([https://www.amazon.com/Kubernetes-Action/dp/1860172773](https://www.amazon.com/Kubernetes-Act([https://www.amazon.com/Kubernetes-Action/dp/180172773](https://www.amazon.com/Kubernetes-Action/dp/1860172773)) – A comprehensive guide to Kubernetes, covering everything from deployment
to scaling.
* **"Kubernetes Up and Running" by David J. Hansen:**
([https://www.amazon.com/Kubernetes-Up-Running/dp/1441034883](https://www.amazon.com/Kubernetes([https://www.amazon.com/Kubernetes-Up-unning/dp/1441034883](https://www.amazon.com/Kubernetes-Up-Running/dp/1441034883)) – A practical guide to getting started with Kubernetes.
**5. Communities & Forums**
* **Kubernetes Slack:** ([https://slack.k8s.io/](https://slack.k8s.io/)) - A great place to
ask questions, get help, and connect with other Kubernetes users.
* **Stack Overflow:**
([https://stackoverflow.com/questions/tagged/kubernetes](https://stackoverflow.com/questions/ta([https://stackoverflow.com/questions/tagged/kubernetes](https://stackoveflow.com/questions/tagged/kubernetes)) - A massive community forum where you can find answers to almost any
Kubernetes-related question.
* **Reddit - r/kubernetes:**
([https://www.reddit.com/r/kubernetes/](https://www.reddit.com/r/kubernetes/)) - A good place
to discuss Kubernetes, ask questions, and stay up-to-date on the latest developments.
**Recommendations Based on Your Experience Level:**
* **Beginner:** Start with the official Kubernetes documentation and Katacoda tutorials.
* **Intermediate:** Take the A Cloud Guru or Udemy Kubernetes for Beginners course.
* **Advanced:** Dive into the Kubernetes documentation, explore advanced topics like
networking, security, and autoscaling, and consider contributing to the Kubernetes community.
**To help me narrow down the best resources for *you*, could you tell me:**
* **What's your current level of experience with cloud computing/DevOps?** (e.g., just
starting out, have some basic experience, familiar with VMs/Containers)
* **What are you hoping to *do* with Kubernetes?** (e.g., deploy web applications, build
microservices, manage containerized applications)
I haven’t been able to test the image capability of the model yet, since my machine is not capable of handling that big of a model, it will barely run the 4B parameter, though I don’t want to crash it.
Overall it looks like a good contender for a go-to model right now if self-hosting for running agentic workflows or just for experiments.