Gemma 4

TAS Score: S3/3 — D4/5 — A3 / T2 (D4 not D5: the hosted option via Google AI Studio / Vertex AI collects usage data and requires a Google account. The self-hosted Apache-2.0 weights scored here carry none of this.)

Family: Gemma 2 / 3 (Gemma Terms) · Gemma 4 (Apache 2.0) This card scores Gemma 4 — the first Gemma under Apache-2.0. Gemma 3 and earlier remain under the custom Gemma Terms of Use (T1) — see the family link.

⚠️ Two-mode tool. Self-hosted open weights: A3/T2. Hosted via Google AI Studio / Vertex AI: A1/T1. This card scores the self-hosted open-weight variant.

Brief Description

Google DeepMind’s fourth-generation open-weight family, released April 2, 2026 — four sizes (2B, 4B, 26B MoE, 31B dense), built on the same research as Gemini 3, natively multimodal, ~256K context. The 2B/4B edge variants run on phones and single-board computers; the 31B dense model runs 4-bit on a 24 GB consumer GPU and ranks near the top of open-model leaderboards.

The landmark change is licensing: Gemma 4 is the first Gemma under OSI-approved Apache 2.0, dropping the custom Gemma Terms that constrained earlier versions. That single change moves the family from T1 to T2.

Architectural Role

Compute/inference layer, full range from edge to workstation. Local alternative to cloud chat and multimodal APIs — now legally clean for commercial deployment.

Technical Autonomy

  • Works without internet (after model download)
  • Stores data locally
  • Does not require external accounts (self-hosted)
  • Allows data export — Apache-2.0, standard HuggingFace format
  • Hosted AI Studio / Vertex option requires a Google account and sends data to Google

Philosophical Assessment (whose.world criteria)

Criterion Status Comments
Pause ✅ Stop inference. Model weights stay on disk.
Exit ✅ Standard model format. Apache-2.0 — no restrictions on use or redistribution.
Recoverability ✅ Re-download from HuggingFace / Kaggle or restore from backup.
Visibility ✅ Apache-2.0. Open weights. Training data not published.
External Dependencies ⚠️ Self-hosted: none. Hosted option: Google infrastructure.

Configuration (Minimal) 

# Edge — 4B on device via Ollama
ollama run gemma4:4b

# Workstation — 31B dense, 4-bit on a 24GB GPU
ollama run gemma4:31b

Alternatives

Alternative Autonomy Notes
Gemma (2 / 3) A3 / T1 Earlier versions — custom Gemma Terms of Use.
Qwen3 A3 / T2 Apache open weights.
Mistral A3 / T2 Apache open-weight family.

Trajectory

Direction: opening

Gemma 4 is one of the clearest opening moves in the catalog: a major lab taking an established restrictive-license family and relicensing its newest generation under OSI Apache 2.0. Google’s own open-source blog frames it as the first Gemma with a truly open license. The remaining caveats are that training data stays unpublished and the hosted path (AI Studio / Vertex) is a separate, data-collecting mode.

Signal assessment:

Signal Status Evidence
License ✅ Apache-2.0 — first Gemma under an OSI-approved license.
Feature gating ✅ All four sizes released as weights, edge to 31B.
Self-hosting ✅ Standard formats. Ollama, vLLM, llama.cpp.
Governance ➖ Google-controlled project, but Apache removes the revisable-terms risk.

Signal key: ✅ opening · ➖ neutral · ⚠️ closing

Sources

This site uses Just the Docs, a documentation theme for Jekyll.