Breaking

Google DeepMind's Gemma 4 reaches 200 million downloads in 2.5 months

June 25, 2026 at 16:50 EDT

Foundation Models
Open Source
AI Agents

Google DeepMind's open-weight model family Gemma 4 has reached 200 million downloads roughly 2.5 months after release. Developers, researchers and startups are fine-tuning and deploying it across a wide range of environments, from wearable robotic arms to enterprise-grade AI security.

June 25, 2026 · Google DeepMind

Gemma 4 hits 200 million downloads in ~2.5 months

Google's open-weight family, built on Gemini 3 research and shipped under Apache 2.0, is being fine-tuned and deployed everywhere — from wearable robotic arms to enterprise security — with download pace accelerating.

200M

downloads in ~2.5 months

256K

token context window

140+

languages supported

Download growth — early June → late June

Same unit (millions of downloads), drawn to scale.

150M

early June

200M

late June

Gemma 4 31B vs the previous Gemma 3 27B

Benchmark scores (%). Orange = Gemma 4 31B · Gray = Gemma 3 27B.

τ2-bench tool use (Retail)

86.4

6.6

AIME 2026 Math

89.2

20.8

LiveCodeBench v6 Coding

80.0

29.1

MMMU Pro Multimodal

76.9

49.7

MMMLU Multilingual Q&A

85.2

67.6

What developers praise

"Very strong for its size"
26B A4B MoE (~4B active): fast, memory-efficient
Practical speeds on Mac Studio & consumer GPUs
12B excels at local coding & image-app generation
Apache 2.0 — no API dependence

Where it falls short

Tool use & agentic functions miss expectations at times
Higher VRAM with long contexts than Qwen-based models
12B feels insufficient for complex tasks

Lineup & reach

Variants span edge/mobile (E2B, E4B) through 12B, a 26B A4B MoE and 31B — runnable fully offline on-device. Available across Model Garden, Hugging Face, Kaggle, Google AI Edge, Ollama and LM Studio.

E2B E4B

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.