Thursday, Oct 8 | --:--
Back to home

Gemma 4 Goes Apache 2.0 — Open Weights Aimed at Agents and the Edge

Google DeepMind’s Gemma 4 family: reasoning/agentic open models from edge (E2B/E4B) to larger MoE/31B-class, multimodal/multilingual (140+ languages), full Apache 2.0 with no commercial restrictions — first clean Apache license in the Gemma line. ‘Most capable open Gemma’ is vendor; treat Arena/agent scores as early-April lab/partner claims.

Times of AI Desk 5 min read Mountain View View as Markdown
Cover illustration for Gemma 4 Goes Apache 2.0 — Open Weights Aimed at Agents and the Edge

Open Gemma releases used to carry custom license friction. Gemma 4’s proprietary frame is the license flip to full Apache 2.0 plus an explicit agentic/edge product story — capability boasts still need independent boards.

Google DeepMind (April 2) released Gemma 4: open models purpose-built for advanced reasoning and agentic workflows, spanning edge (phones, Raspberry Pi-class) to server. Variants include efficient E2B/E4B and larger 26B MoE / 31B-class. Multimodal, multilingual (140+ languages), long context for multi-step planning, autonomous action, code, audio-visual tasks. Same research lineage as Gemini (company). Apache 2.0 — modification, commercial use, redistribution without prior custom restrictions. Hugging Face day-one support. Google AI Edge companion post: bring agentic skills on-device without constant cloud round-trips.

Claims vs checks

Family sizes, Apache 2.0, and edge positioning are Google primary (The Keyword / AI Edge). “Most capable open models” / frontier-level reasoning and high Arena-style scores as of early April are vendor. Independent Artificial Analysis or Arena Elo should be checked separately before freezing a ranking.

Limits

  • Open weights ≠ free of Acceptable Use / policy overlays in hosted products.
  • Agentic quality on-device depends on quantization and hardware — not guaranteed by the blog.
  • Early-April partner benches can be selective.

Sources

Prior Coverage

Earlier Times of AI reporting on this thread.

Scroll to continue reading