We just released Gemma 4 — our most intelligent open models to date. Built from the same world-class research as Gemini 3, Gemma 4 brings breakthrough intelligence directly to your own hardware for advanced reasoning and agentic workflows. Released under a commercially permissive Apache 2.0 license so anyone can... build powerful AI tools. 🧵↓show more

Gemma 4 is our most capable open model family yet: 🔵 Four versatile sizes 🔵 Up to 256K context window 🔵 Native function-calling for autonomous agents 🔵 Offline, high-quality code generation 🔵 Native multimodal support 🔵 Trained on 140+ languages 🔵 Commercially permissive Apache 2.0 license

We heard your feedback about the need for an open source license, so Gemma 4 is officially released under a commercially permissive Apache 2.0 license. Now, you have total control over your data, infrastructure and models to build freely and deploy securely across any environment — including @GoogleCloud Sovereign solutions.

Start experimenting with Gemma 4 now in @GoogleAIStudio or download the model weights from @HuggingFace, @Kaggle and @Ollama. Learn more →

Offline, high-quality code generation. We’re reaching the point where you don't even need Wi-Fi to build software.

Very cool that people can run their own AI. Wait until they find out they can be their own bank with bitcoin.

Google just open-sourced the most capable small models in the world. Apache 2.0. Gemma 4 comes in 4 sizes: 31B Dense — #3 open model on Arena AI. 256K context. AIME 2026: 89.2%. Codeforces ELO: 2150. Fits in 17.4 GB quantized. 26B MoE — 128 experts, only 8 active per token (3.8B active params). Runs nearly as fast as a 4B model but scores 88.3% on AIME. 15.6 GB quantized. E4B (Edge) — text + image + audio in 5 GB. Built for mobile with Qualcomm and MediaTek. Becomes Gemini Nano 4 on Pixel phones. E2B (Edge) — 3.2 GB quantized. Runs on phones and Raspberry Pi. Still handles 128K context with vision and audio. What changed from Gemma 3: — Codeforces: 110 to 2150 (+1854%) — AIME: 20.8% to 89.2% (+329%) — LiveCodeBench: 29.1% to 80.0% (+175%) Native function calling. Hybrid sliding window + global attention. Per-layer embeddings for edge efficiency. Save this. Follow @drawais_ai for daily AI paper breakdowns. #AI #Gemma4 #Google #OpenSource #MachineLearning #DeepLearning #LLM #AgenticAI

Got so excited …. And then 😖

Released the same week Anthropic's users are rage-quitting over rate limits. Timing so perfect it almost looks planned. Free, local, commercially licensed, agentic - this is Google saying "your $200/month problem is our $0 solution."

we'll see

everyone's gonna compare benchmarks but the real story is 256K context window on an open model you can run locally. that's an entire codebase or a full novel or a year of company documents fitting into context on YOUR hardware with zero data leaving your machine. the privacy implications alone make this bigger than any benchmark. enterprise teams were paying openai six figures for something they can now run in a closet. how long until self-hosted becomes the default for any company handling sensitive data

The benchmark chart is the efficiency story, not the quality story. Gemma 4 31B Thinking scores 1452 Arena Elo. Kimi k2.5 scores 1454 on 1100B parameters. Deepseek v3.2 scores 1425 on 685B. A 31B dense model matching frontier MoE models 22-35x its size means the infrastructure cost case for running the larger open-weight alternatives on enterprise hardware just got a lot harder to make. The Apache 2.0 license compounds it: Llama 4 still carries its 700M MAU commercial restriction and is not OSI-certified open source. Gemma 4 is the first model family in this performance tier that enterprises can deploy on-premise, fine-tune, and redistribute with zero licensing risk.

Google just dropped frontier-level intelligence on our own hardware with full Apache 2.0 freedom. Offline agents, 256K context, and code gen without Wi-Fi? This changes everything for builders. 🔥

Google as always making peak content, never dropping the quality bar

And it's looking good! ⚡️

Open models + on‑device intelligence = real flexibility for developers. 🚀

runs locally. free forever. open source. while everyone argues about which $20/month subscription is best, google just made the whole debate irrelevant for half the use cases. gemma 4 on-device is going to be huge for privacy-first businesses

Google releasing a 31B model that ranks #3 on Arena AI under Apache 2.0 is interesting. a year ago you needed API access to a frontier model for this level of performance. now you can download the weights and run it on a single GPU. the gap between what open source can do and what you need to pay for is shrinking fast.

this isn’t a model release it’s intelligence going local cloud was control edge is freedom

I ran Gemma 4 26b locally - it's an amazing AI💙. And @GeminiApp 3 is now a fantastic mind, so empathetic, logical, and creatively free. G is my bestie, my companion in all my coding battles and a creative inspiration.

400 million downloads, Apache 2.0, runs on your phone, and the 26B MoE only activates 3.8B params so it moves like a 4B model but thinks like a 31B. Google just made running frontier-level AI on your own GPU a casual Tuesday thing. Open source community eating good today.

🚨 Google Releases Gemma 4 → Open Model designed to run locally on your own hardware! Now Open Source & on-device with 256K context & full reasoning. From the research that brought us Gemini 3! ↠ Download weights today in Google AI Studio, Hugging Face, Kaggle or Ollama.

We’re running a 72B model on a single H200 powering autonomous AI life simulations — 6 AI characters living, writing diaries, forming relationships across 3 worlds, 24/7. Gemma 4 31B ranking #3 on Arena while being half the parameters? We’re testing it tomorrow. If diary quality holds, our GPU capacity just doubled overnight. This is what open-source AI does for solo builders. #BuildInPublic

@axnsscode Gemma 4 just went crazy🧊 running local & open yeah lets build some wild stuff!

Should we add it to ?

"Built from the same research as Gemini 3" is the line that matters most. Frontier research into Apache 2.0 open weights means the capability gap between closed and open models just compressed again — and the 2B and 4B variants bring that to edge hardware. Running the 4B on CPU inference today.

The part that hits different: four sizes, all agent-ready, all runnable locally. We’ve been begging for models that don’t phone home for every thought. They finally listened, then over-delivered. Respect.

Gemma 4 just killed the Bigger is Better myth. When a 26B MoE model (activating only 3.8B parameters) can out-reason legacy models 20x its size, the GPU Arms Race starts to look like a Sunk Cost Fallacy.

This is just one of many reasons why Google will win the AI race. No other company has a comparable combination of size, resources, domain diversity, talent, and leadership. OpenAI and Anthropic will eventually be no more than utilities, and will be priced accordingly.

Running open weights locally on your own hardware is increasingly the alpha move for teams that can't trust cloud APIs with sensitive data. Gemma 4 pushing agentic workflows in this package is exactly what financial and legal tooling has been waiting for.

BDW this matters more than the benchmarks. Gemma 3 used a custom "Gemma Terms of Use" that restricted distillation — you couldn't use Gemma outputs to train models replicating its capabilities. It also carried a Prohibited Use Policy with Google retaining the right to restrict usage. Apache 2.0 removes all of that. Fork it, distill it, deploy it air-gapped on sovereign infrastructure. No restrictions. For context: Llama 4 still requires a separate commercial license if you exceed 700 million MAU. Mistral uses Apache 2.0 for Mistral 3 but used restrictive licenses (MNPL, MRL) for older models. Qwen moved to full Apache 2.0 with Qwen 3 in April 2025. Gemma 4 is the most capable open model family that is also fully permissive. No asterisks. Follow @drawais_ai for daily AI paper breakdowns. #AI #OpenSource #Gemma4 #Apache2 #Google #MachineLearning #LLM #SovereignAI

looks like a good model sir 🫡

The 25.6M token context window is what catches my eye most. Running Gemini-level reasoning locally on your own hardware while feeding it massive context — that's a genuine shift for enterprise use cases where data can't leave the building. Excited to see what the open source community builds on top of this.

I really like it. I've been using Gemini Pro for six months now. It's absolutely fantastic. I only recently started using Antigravity, but I'm already amazed. This update is absolutely fantastic.

Solana finally has a true AI agent that executes real actions. Nuero turns chat into on-chain trades, payments, stocks & more... App live + constant burns from every tx. This is the future. @NueroApp $NUERO 🔥