Gemma

TAS Score: S3/3 — D4/5 — A3 / T1 (T1 not T2: Gemma 2 and 3 ship under the custom “Gemma Terms of Use” — a source-available license with a prohibited-use policy that Google may revise and enforce remotely. Not OSI-approved. Weights are downloadable and self-hostable, so autonomy stays A3; license transparency is capped at T1.) (D4 not D5: the Gemma Terms are revisable by Google and include a remote-restriction clause — a standing diagnostic concern.)

Family: Gemma 2 / 3 (Gemma Terms) · Gemma 4 (Apache 2.0) This card scores Gemma 2 and 3, under the custom Gemma Terms of Use. Gemma 4 moved to Apache-2.0 — a separate object, see the family link.

Brief Description

Google DeepMind’s open-weight model family, built on the same research as Gemini. Gemma 2 (June 2024) and Gemma 3 (March 2025) span roughly 2B to 27B parameters, run on consumer hardware, and are widely supported (Ollama, vLLM, llama.cpp). The weights are downloadable and self-hostable — but the license is the custom Gemma Terms of Use, a source-available license with a prohibited-use policy that Google can update and enforce remotely. That is not OSI-approved, which is why these versions land at T1 despite running fully offline.

Gemma 4 (April 2026) changed this — see the family card. Gemma 3 and everything earlier remain under the original terms.

Architectural Role

Compute/inference layer: general-purpose open-weight LLM run on your own hardware. Local alternative to cloud chat APIs on modest hardware.

Technical Autonomy

  • Works without internet (after model download)
  • Stores data locally
  • Does not require external accounts (self-hosted)
  • Allows data export — weights downloadable (Gemma Terms of Use; prohibited-use policy applies)
  • Subject to Google’s revisable usage policy

Philosophical Assessment (whose.world criteria)

Criterion Status Comments
Pause ✅ Stop inference. Model weights stay on disk.
Exit ✅ Standard model format, downloadable.
Recoverability ✅ Re-download from HuggingFace / Kaggle or restore from backup.
Visibility ⚠️ Open weights, but custom Gemma Terms (non-OSI, revisable) → T1. Training data not published.
External Dependencies ⚠️ Self-hosted: none at runtime. License: Google’s revisable terms.

Configuration (Minimal) 

# Consumer hardware via Ollama
ollama run gemma2      # 9B / 27B, Gemma Terms of Use
ollama run gemma3      # multimodal, Gemma Terms of Use

Alternatives

Alternative Autonomy Notes
Gemma 4 A3 / T2 Same family — Apache-2.0, OSI-clean.
Qwen3 A3 / T2 Apache open weights, no custom terms.
DeepSeek-R1-Distill-Llama A3 / T1 Same “open weights, non-OSI license” shape.

Trajectory

Direction: opening

The story here is motion. Gemma launched in 2024 under a restrictive custom license and stayed there through Gemma 3 — but Gemma 4 (April 2026) became the first Gemma under OSI-approved Apache 2.0. These earlier versions remain T1, yet the family’s direction is one of the few genuinely opening trajectories in the catalog: a major lab relaxing its license rather than tightening it.

Signal assessment:

Signal Status Evidence
License ⚠️ Gemma Terms of Use — custom, non-OSI, revisable. (Gemma 4 fixed this.)
Feature gating ➖ All sizes released as weights; no capability withheld.
Self-hosting ✅ Standard formats. Ollama, vLLM, llama.cpp.
Governance ⚠️ Google may update the usage policy and restrict use remotely.

Signal key: ✅ opening · ➖ neutral · ⚠️ closing

Sources

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