Gemma 4 — the license change is a bigger story than the model

The take

Google released Gemma 4 on April 2 — 31B dense and 26B MoE (3.8B active per token) variants, 256K context native, vision + audio multimodal. Reached Arena top 5 in its first week of community evaluation. Standard release-week noise.

The license change is the bigger story. Gemma 1, 2, and 3 all shipped under Google's custom 'Gemma Terms' — a quasi-permissive license that Google pitched as 'effectively open' but contained enough restrictions to make commercial redistribution legally fraught. Lots of builders quietly avoided Gemma in production for that reason.

Gemma 4 ships under Apache 2.0. That's a clean, OSI-approved, commercially-friendly license — the same one Qwen and Mistral use. Google explicitly framed this as 'expanding the Gemmaverse with Apache 2.0' in its own announcement. For product builders, this removes a real friction.

Practical: Gemma 4 31B Dense fits 24 GB at Q4 with room. 26B MoE (3.8B active) is the speed pick for 16–24 GB cards. Both ship native Ollama tags (`gemma4:31b`, `gemma4:26b`). Both are reasonable picks for chat.high or docs.high — though Qwen 3.6-27B (dense) and 3.6-35B-A3B (MoE) currently lead those slots on this site for benchmark reasons, Gemma 4 is the safe Apache-2.0 alternative when license clarity outweighs marginal benchmark gains.

Where it sits: chat.top (#2), docs.top (#3), chat.high (#2). The dense 31B variant is the calmer long-context behavior pick; the MoE 26B is the faster general-purpose option.

Where this fits

Models: Gemma 4 (31B dense + 26B A4B MoE + 12B multimodal) · Qwen 3.6-27B · Qwen 3.6-35B-A3B

Hardware: NVIDIA RTX 5090 · NVIDIA RTX 4090 · NVIDIA RTX 3090 (used, single) · AMD Radeon RX 7900 XTX · AMD Radeon RX 9070 XT

Sources

• Google — Gemma 4 with Apache 2.0
• VentureBeat — Gemma 4 Apache 2.0 license matters
• HuggingFace — Gemma 4 release

Next step