Meet Gemma 4! Purpose-built for advanced reasoning and agentic workflows on the hardware you own, and released under an Apache 2.0 license. We listened to invaluable community feedback in developing these models. Here is what makes Gemma 4 our most capable open models yet: 👇

Apr 2, 2026 · 4:07 PM UTC

164
812
7,035
630,238
Learn more in the blog blog.google/innovation-and-a…

Gemma 4: Byte for byte, the most capable open models

Gemma 4: our most intelligent open models to date, purpose-built for advanced reasoning and agentic workflows.

blog.google
12
21
215
50,071
Sort replies: Relevant Recent Liked
Replying to @googlegemma
Nice work
82
62
890
36,367
Replying to @googlegemma
I’ve already gotten out of bed, started downloading, and I’m getting ready to study and analyze. 💪
1
21
1,616
Replying to @googlegemma
Gemma 4 is a strong step for AI at the edge. 💡
8
609
Replying to @googlegemma
Your predecessor was cool to analyze!
Ramp Labs
@RampLabs
Apr 2
Article

How we built Steer, our interpretability playground

We built a system that steers an LLM toward specific concepts without retraining, and in the process learned something concrete about where meaning lives in different model architectures. Steer a
4
235
Replying to @googlegemma
4
1,448
Replying to @googlegemma
when is Gemma 5 dropped?
1
953
Replying to @googlegemma
native function-calling changes how we build agents. instead of using "hacky" prompts to get an ai to use a tool, gemma 4 has built-in support for structured json, meaning your local ai can now reliably control your apps and apis without making mistakes.
1
642
Replying to @googlegemma
Let's go
1
305
Replying to @googlegemma
LFG Google
79
Replying to @googlegemma
Gonna play around with this.
33
Replying to @googlegemma
I want to put Gemma 4 to our VPN app, how about that?
69
Replying to @googlegemma
I don't think the rating is true, I really like Gemma-4 it's way much better but maybe not better than Qwen-3.5
344
Replying to @googlegemma
Nice
179
Replying to @googlegemma
31B DENSE Wouldve loved to see a bigger model though
531
Replying to @googlegemma
already meet this AI i'm excited right now to see on Gemma 4
33
Replying to @googlegemma
ME News
@MetaEraCN
Apr 3
Google 发布迄今为止最智能的开放模型——Gemma 4 @Google 今日正式推出Gemma 4系列开放权重模型，基于Gemini 3同等世界级研究打造，成为其迄今为止最智能的开放模型。 核心亮点： 1.四大尺寸可选，最高支持256K超长上下文 2.原生函数调用 + 多模态能力 3.支持140+语言，商业友好Apache 2.0许可 4.专为代理工作流、离线代码生成、边缘设备深度优化 可在本地硬件直接运行
40
Replying to @googlegemma
기대된드아~~
47
Replying to @googlegemma
Local model that fits in a single GPU on par with glm-5 would be insane. Can't wait to test.
29
2,193
Replying to @googlegemma
Apache 2.0 on local hardware is exactly what the developer community needed. Too many "open" models still had strings attached. Running advanced reasoning locally without licensing headaches changes everything for production deployments.
22
1,824
Replying to @googlegemma
outside of the Arena elo score, we compared the gemma models head to head against qwen on common benchmarks, and found remarkably different results: maniac.ai/blog/qwen-3-5-vs-g…
1
1
13
2,017
Replying to @googlegemma
Apache 2.0?! Amazing. Performance looks incredible too. Congrats to the whole team on this one, you've really cooked!
10
1,101
Replying to @googlegemma
the intelligence density of those small models is insane, gpt oss 120b feels so behind now
7
1,240
Replying to @googlegemma
What the hell is that
3
10
1,342
Replying to @googlegemma
Tested Gemma4 31B - after short tests, many things seems worse than in Gemma3 27B, lots of "glitches" - infinite repeating loops. On my 5090 Im abe to run it only with max 12000 token context (lm studio community Q4_KM). With Qwen3.5 27B I can run full 260000 context (also Q4 KM)
1
7
1,041
Replying to @googlegemma
Running agents locally with Gemma 4 sounds great. But I've seen this pattern before - works great solo, falls apart in multi-step chains. Testing local agentic stacks is still unsolved. Curious what @GoogleDeepMind's eval story looks like.
1
4
1,329
Replying to @googlegemma
Demis Hassabis needs only a DGX spark to compete with the other AI labs, Elon Musk needs ban entire space station with a base on the moon. The other labs need an entire data centre. Levels!!!🙂‍↔️
1
2
109
Replying to @googlegemma
3
654
Replying to @googlegemma
well that's a crazy start waiting for the 4b and 2b variants and i won't say no for a tech report
6
890
Load more