Gemma 4: open multimodal models under Apache 2.0
Google DeepMind's Gemma 4 family is Apache 2.0, accepts images or video plus text (and audio for the small models), and adds Per-Layer Embeddings and a shared KV cache for efficiency.
What happened
Gemma 4 was released, with support across Transformers, llama.cpp, MLX, WebGPU/transformers.js, Mistral.rs and fine-tuning with TRL and Unsloth.
The technical details
From the Hugging Face post:
- Apache 2.0 licence
- all models take image (or video) and text inputs. The small E2B and E4B variants (and, per the post, a 12B Unified model) also take audio.
- an image encoder like Gemma 3’s, with two improvements: variable aspect ratios and a configurable number of image tokens (trading speed and memory against quality)
- architecture features the post highlights: Per-Layer Embeddings (PLE) and a shared KV cache
- the authors say the models were so good out of the box that they struggled to find good fine-tuning examples
Why it matters
Gemma started in February 2024 as text-only 2B/7B models. Two years later it’s multimodal with audio, efficient enough for devices, and permissively licensed.
#llm#multimodal#on-device#open-weights
Sources
- Welcome Gemma 4: Frontier multimodal intelligence on device — Hugging Face
This is a summary written from the linked public sources. Figures and numbers are as reported there.