May 28 Update: Winner announcement delayed until June 18: read the update here
Local AI is having a moment, and we want you to be part of it!
...
For further actions, you may consider blocking this person and/or reporting abuse
Displaying a subset of the total comments. Please sign in to view all comments on this post.
This challenge is either a make or break for me! Will probably do both, but we will see.
@jess If you are doing both prompts, would it be appropriate to do in one post or do you prefer to be separate post. I say this because I was thinking about writing about Gemma 4, but also showcase what I built (not going to spoil it yet).
Regardless, hope to see what the community builts/writes :D
Hey @francistrdev! I would keep those as two separate submissions. The content should be different though, please don't just duplicate two posts.
Sounds good! Thanks! I also sent you an email on other stuff! Noticed you might be busy lol. lmk!
Interesting perspective. Curious how others are handling this.”
This one is going to be great. I’m most excited about the IOT related use cases, but it all seems cool.
I explored Gemma 4 through the lens of local AI ownership— how moving from cloud-only APIs to capable local models changes the experience for users. One of the most interesting parts for me was intentionally choosing the E4B model instead of defaulting to the largest variant, because accessibility matters just as much as raw capability. Feel free to check out my submission 😇
Excited to read what everyone else is building and writing!
love this, definitely joining in 😄
This is a yummy challenge!
Hope to see a lot of really useful projects, and at the same time some useless but still cool projects!
Love it 😍
Thanks to Google and the Dev team!
This is a cool challenge. Gemma 4 should lead to some really interesting projects.
Rock & roll
I want to.. I need to find something I could do with my potato laptop. it's hot even to compile cpp with 20 lines still waiting for +5 mins.
ahh, theres solution <3 thank you google & ollama for the cloud <3 amazing!
Nice! I've already got some good testing with it in my agents memory.
This is exciting. definitely joining this challenge. 🤙
Nice challenge!!
thank you for this amazing oppurtunity!❤️👊🏻
This is my submission : dev.to/shogun444/gemma-4-on-16gb-r...
Right on time! I am building with it now, super excited for this opportunity. 🥳
can i download it from ollama?
Yh apparently you can, it's quite big though
ollama.com/library/gemma4/tags
Ollama has it available on their cloud platform. I believe yyou get a pretty decent amount on the free plan for gemma. If you run Opwe WebUI or equiv google ai studio has a pretty generous free tier for gemma 4 models.
Just finished posting about the Gemma 4 family- feel free to go through it.
Now I am on to the next venture, the "Build with Gemma 4" project as it is mentioned over here, and hope it stands out as well. And I also can't wait to go through the projects created by the other builders. The challenge is great. Best of luck everyone.
Excited for this one. I've been building an AI sales chatbot for Arabic-speaking merchants (Provia), and "intentional model selection" is a criterion I take personally — most multilingual benchmarks barely scratch Arabic generation quality in production contexts.
Planning to submit to the Write track: a head-to-head on Arabic e-commerce conversations, same prompts, same product catalog, controlling for everything except the model. Curious whether 140 languages means fluent or just supported.
Hoping to bring something useful to the community. Good luck everyone 🚀
Just published my submission for the Write About Gemma 4 track.
I built an ICS Tabletop Exercise Simulator -- a single Gemma 4 26B MoE model simultaneously simulating six Incident Command System positions so Emergency Operations Managers can run realistic training exercises alone, without coordinating a room full of people.
The write-up covers the architecture, a real token loop issue I hit with extended reasoning on complex multi-constraint prompts (and the fix), an honest look at what happened with RAG retrieval quality, and why the 26B MoE specifically was the right model for this workload.
[dev.to/kkierii/i-used-gemma-4-to-s...]
Hey @jess @ben @jonmarkgo . I got a quick question on top of this since this is important for the Writing Portion of this challenge.
My post for the writing portion has other people involve. The way it is structured the writing submission is that I talk to @codingwithjiro, @javz, and @konark_13 about their opinion on Gemma 4. They sent me their report and I mention them onto a single post and paraphrasing their experiences. Note that I am the main writer on the submission and they do not have access to edit the submission. They only sent me the report about their Gemma 4 and I am paraphrasing it into a single post. I am crediting them on the post that I worked with them and communicate them about Gemma 4.
I was wondering if this counts as a valid individual submission because of this rule:
If I am writing the piece that mentions other people's experience using Gemma 4, would that still count as an "individual submission" or where do you draw the line? Is this fine?
That is fine with us as long are your guests are aligned that this is your individual submission.
Just submitted my entry. I built a local computer vision pipeline on a Raspberry Pi 5 using Gemma 4's native bounding box output. Replaced my entire YOLO + OpenCV setup with 50 lines of code. The zero-shot detection capability is honestly what sold me on this model, no retraining needed for new object categories. Excited to see what others are building.
The model selection framing is what I find most interesting about this challenge. Most AI challenges just say "build something with X" and leave it there, but the emphasis on why you chose the model you did as a judging criterion is a different ask entirely. It pushes you to actually think about the tradeoffs between the E2B, E4B, and the 31B dense not just grab whichever one produces the best demo output.
@ben point about IoT use cases caught my attention too. The fact that the E2B model can run on a Raspberry Pi 5 opens up a genuinely interesting design space local inference at the edge, no cloud dependency, no latency from a round trip. That's a different category of application than most AI challenges enable, and I'm curious to see what people build in that direction.
I'm currently in the middle of GSoC so bandwidth is tight, but the write track feels accessible even with a constrained schedule a comparison piece breaking down when you'd actually reach for each of the three model variants would be genuinely useful to the community and doesn't require spinning up a full project. The OpenRouter free tier being available for the 31B is also a nice touch for people who want to experiment without setting up local hardware first.
One quick clarification question for the team: for the write track, is a post that includes a small working code example and a walkthrough treated as a write submission or does it cross into build territory? Trying to understand where that line sits before I decide on an angle.
Thanks for sharing this challenge! I'm interested in participating. Could you clarify:
· What are the judging criteria (e.g., creativity, technical complexity, real-world use)?
· Are there any restrictions on which Gemma 4 model size we can use?
· Is fine-tuning allowed, or only prompt engineering?
· Where should we submit the final project (GitHub + Dev.to post)?
Appreciate the $3K prize pool — excited to build something useful with Gemma 4!
Eu gostaria de fazer parte do projeto.
Submitted for Gemma Build Challenge -
Gemma.Witness: offline, signed evidence capture for field journalists. Audio plus photos go through Gemma 4 locally, out comes a .witness bundle anyone can verify by dragging it into a static HTML page. No server, no network, no trust required.
dev.to/moonrunnerkc/gemmawitness-o...
Interesting challenge.
What I’m curious about is how people evaluate these models beyond demos now.
A lot of projects look impressive in short workflows, but the real separation starts showing up with:
That’s where things usually get much harder.
Can I use Gemma 4 with React Native? Am thinking of building an app with rn-executorch but am not sure it Gemma is supported
Excited to participate in the Gemma 4 Challenge!
I've built an open-source project called PromptGuard, a local AI Privacy Firewall that sanitizes sensitive information before prompts are sent to AI tools.
Gemma 4 Challenge: Build With Gemma 4 Submission
🛡️ PromptGuard: I Built a Local AI Privacy Firewall That Sanitizes Your Prompts Before They Leave Your Machine
I've been running Gemma and other small LLMs locally for recent side projects, and the ability to prototype without API rate limits has been a game changer for experimentation. The 4B parameter sweet spot seems to hit the right balance of capability and consumer hardware accessibility. Are participants required to use specific quantization levels, or is there flexibility to optimize for different GPU VRAM constraints?
Just submitted to the Write category. Tested Gemma 4 26B MoE and 31B Dense against GPT-4o-mini and GPT-4o as the customer-facing reply in my Arabic e-commerce chat router — six real customer scenarios, real catalog.
The surprise wasn't hallucination. Both Gemma variants in Round 1 stalled or hedged instead of fabricating. So I added three Gemma-only rules — Arabic-first frame, temperature 0.3, max_tokens 400. Same intervention, opposite directions: the 26B MoE flipped toward grounded catalog answers; the 31B Dense flipped toward false-negative refusals — "we don't have that" with the answer sitting in its context.
Architecture > size, at least for prompts that need "check first, refuse on absence" sequencing. The a4b suffix on gemma-4-26b-a4b-it ended up mattering more than the parameter count.
dev.to/alimafana/i-added-three-rul...
Thanks for sharing this challenge! I'm interested in participating. Could you clarify:
· What are the judging criteria (e.g., creativity, technical complexity, real-world use)?
· Are there any restrictions on which Gemma 4 model size we can use?
· Is fine-tuning allowed, or only prompt engineering?
· Where should we submit the final project (GitHub + Dev.to post)?
Appreciate the $3K prize pool — excited to build something useful with Gemma 4!"
It took almost a week of research, reading model cards, reviewing benchmark papers, validating community findings, and refining arguments to complete this article.
Because of the challenge deadline, I decided to publish the current version today. While I'm happy with the analysis, I see this as Version 1 rather than the final word on the topic.
My goal after the competition results are announced is to expand the article further with additional perspectives, deeper experimental evaluations, real-world deployment case studies, multilingual testing insights, and broader comparisons across the open-model ecosystem.
For now, I'm grateful to everyone who takes the time to read, critique, challenge, or contribute ideas. Thoughtful discussion is often more valuable than agreement.
Gemma 4 Challenge: Write about Gemma 4 Submission
Choosing the Right Gemma 4 Model Matters More Than Choosing the Best One
Thank you for being part of the journey.
Just dropped a first-principles breakdown of Gemma 4 — from how Per-Layer Embeddings let a 1.5 GB model do competition mathematics, to exactly which model belongs on which hardware. No fluff, no press release rehash. Would love to know what you think. LINK
In this challenge, what are the standards for selecting the 10 winners?
For example, should the project focus more on being useful for users or on having a beautiful design?
They have mentioned in the - "Judging Criteria" section.
Wow, excellent opportunity
Just joined this platform today and I landed here after login in
I haven
t see this challenge, so Ill just participate.....and im still have no idea what to build or write. -_-_
Are you supposed to build something with the help of Gemma 4, or actually include AI features like chat which relies on Gemma 4?
Can we build something with the assistance of Gemma 4 but not include any such feature which relies on a LLM?
@jess Sorry to ping you! I feel I may have mis-understood the requirements.
Thanks for the Invitation, Perfect Timing . . .
En cuanto al reto se puede presentar el archivo github para que lo valoren o es lo escribir el articulo y nada más
Given that I had recently done some quick benchmarking with Gemma 4 on my Lennova Legion Pro. (64 GB of Ram RTX 5080 card with ~16GB VRAM. I decided to take my notes, and turn them into an experience report.
Some of my findings were surprising.
mine is here:
dev.to/timothy_western_ed7594e0a/p...
This is so great
Cool 😋
This is great....
I will definitely be part of these 👌
Ai 不能太信任
Im business started to Pakistan and work former employee rice whole seller delar
This one is going to be great. I’m e cases, but it all seems cool.🫠
Looks like a fun challenge 👀
Really curious to see what people build with local multimodal models.
Interesting perspective. Curious how others are handling this.”
Hii i am deepak !
I am new to this community so can anyone tell me how to register for this Gemma 4 Challenge:
@jess hii I am Deepak and I am 20yrs old i completed my project and submitted... Can you check that my project was submitted correctly for me
I love this!!
Can I fine tune?
This will be fun. I'll definitely join!
I will say the Gemma api has been getting a lot of errors recently so be careful and don't forget your rate limits.
Gemma 4's "128K context window" seems interesting, but I'm curious about its real-world scalability challenges. Has anyone tried deploying their 31B parameter dense model on AWS? For those working on the building side, I've been using prachub.com for system design mocks. Their follow-up questions on latency are really similar to what my interviewers have asked. What are you guys planning to build with Gemma 4?
On my way ❤️🔥
Excited to submit my entry for the Gemma 4 Challenge!
“GemmaOps Edge” focuses on reasoning over real network signals — not just basic LLM inference — combining alarms, topology, and history to deliver intelligent RCA insights, fully on-device.
Looking forward to exploring other amazing submissions here!
dev.to/pravdexter/gemmaops-edge-fr...
@jess i just seen this challenge today.
I have been building my new AI saas, narasi.ai since 2 months ago, and gemma 4 (26b & 31) is the backbone of this project.
So, do i qualify to submit this project for this challenge? Its a whole production app (with payment gateway) that is just released yesterday
Great!
uhh
Just submitted to Build With Gemma 4: L.E.N.S. — a voice-first photography coach for blind and low-vision artisans listing handmade work online.
The maker can't verify the photo the way they verify the craft by touch. Gemma 4 E4B (Ollama, local) reads the actual product image and returns structured JSON → one spoken fix → compare two shots → copy-ready listing text.
I picked E4B on purpose: multimodal enough to trust, small enough to run offline on a maker's laptop — privacy isn't a feature here, it's the independence mechanism.
No-install demo + walkthrough in the post. Would love feedback if you try it.
Post: dev.to/prasadt1/lens-a-private-pho...
The GIL release on I/O is the part most people miss. Your threading example hitting the same speed as single-threaded is the classic tell — 8 threads all fighting over one lock while the actual work (parsing, dict lookups, string ops) is pure bytecode.
The real question nobody asks: for your markdown parsing workload, would multiprocessing with shared memory have been faster than threads? The pickle serialization cost on a few hundred thousand small objects can eat the gains, but if you're just pushing text through a pipeline, ProcessPoolExecutor sometimes beats ThreadPoolExecutor by 2-3x on CPython.
Also: did you try PyPy? The JIT can eliminate some of the GIL pressure for tight loops, though the warmup time hurts one-off scripts.
The "errors are values" philosophy is clean in theory but the wrapping chains get brutal in real codebases. Your Pattern 4 (sentinel errors) is where most Go teams I've seen eventually land — fmt.Errorf with %w is fine until you have 5 layers of wraps and errors.Is starts feeling like instanceof hell.
The Rust comparison at the end is fair. What Go's error model actually wins on is readability at the call site — you can see exactly what can go wrong without jumping to a type definition. Rust's ? operator is more ergonomic but it hides the error path unless you're reading carefully.
One thing this piece doesn't cover: error context in concurrent code. When 10 goroutines return wrapped errors, how do you reconstruct which request caused which failure? multierror or errgroup.Group? That's where Go's model gets messy fast.
The offline constraint forces you to think about where the "AI layer" actually adds value vs where it's just cargo-culted in. For code completion, sure — local models are fine. But for reasoning-heavy tasks like code review or architecture decisions, the latency and capability gap between local and frontier models matters more than people admit.
One thing worth noting on the security angle: air-gapped means you also lose model updates, CVE patches for the inference stack, and any telemetry that catches model-level vulnerabilities. The tradeoffs aren't just about capability.
Curious what your context window strategy is with the 8B model — do you split long files into chunks manually, or does the tooling handle it? That's usually where local setups break down for real engineering work.
The Bedrock AgentCore config-driven approach is the right abstraction for this use case. YAML-driven agent definitions map cleanly to state machines, and the "deploy from config" model is how most agent frameworks should have started instead of requiring code for everything.
The "roll your own" vs OpenClaw tradeoff is real though — OpenClaw handles the local execution, event loop, and persistence that sounds trivial until you're building it yourself. Storage, scheduling, error recovery, memory management... the 20% that takes 80% of the effort.
For the "control it" requirement — what does that look like in practice? For edge deployments (robotics, IoT), you'd want the agent's memory to survive restarts without losing context. Do you persist the conversation history between scheduled runs, or does it reset each cycle?
Applied ❣️
Gemma 4 Challenge: Build With Gemma 4 Submission
🧞♂️Transform unstructured PDFs Job Offers into a dataset w. gemma4:2b
I am definitely getting in on this. Not even "in it to win it". I just can't wait to see all the great things people come up with.
What stood out to me was how much time QA teams actually spend maintaining locators instead of writing tests. The comparison between traditional Appium Inspector workflows and Vision AI-based testing was explained really well. I also liked that the blog didn’t completely dismiss Appium Inspector and clearly showed where it is still useful. Interesting read on how mobile automation is evolving beyond heavy XPath dependency.
Super excited for this challenge! 🚀 The Gemma 4 model family looks incredibly versatile especially love that the 2B/4B models can run on edge devices like Raspberry Pi and Pixel phones. Local AI is definitely the future. Free
Building with Gemma has been a game-changer for lightweight AI apps. Excited to be part of the #gemmachallenge! What do you all think of the implementation?
I didn`t see the Gemma 4 Submission Template when I clicked the first link
Here's my submission:
dev.to/raslanove/insults-cutlasses...
Thank you for this opportunity!
Submitted my entry ✨
dev.to/samirmishra27/i-built-proof...
Happy to take part in this amazing opportunity!
I'd love to rank up on this tournament
Interesting perspective. Curious how others are handling this.”
I love what is cooking here
I was actually making a game, does it count?
¿que es gemma4?