Gemma

Every Gemma release before this one shipped under the custom Gemma Terms of Use, which grants broad commercial rights but incorporates a Prohibited Use Policy by reference, includes a clause letting Google remotely restrict access it believes violates the agreement, and requires redistributors to pass the terms downstream (https://ai.google.dev/gemma/terms). That page is still live and still governs Gemma 3 and the specialized variants. Gemma 4, released April 2, 2026, switched to plain Apache 2.0 — confirmed on its official model card (https://ai.google.dev/gemma/docs/core/model_card_4), on Google's Open Source blog announcement (https://opensource.googleblog.com/2026/03/gemma-4-expanding-the-gemmaverse-with-apache-20.html), and independently in the license field of every Gemma 4 repo on Hugging Face's API (checked directly: google/gemma-4-31B-it, gemma-4-26B-A4B-it, gemma-4-12B-it, gemma-4-E4B-it, gemma-4-E2B-it all report license:apache-2.0). Apache 2.0 drops the acceptable-use flow-down and the unilateral kill-switch language — a meaningful difference if you were avoiding Gemma or Llama for exactly that reason. It does not grant trademark rights to the Gemma name.

Gemma 4 ships in five sizes — E2B (2.3B effective params), E4B (4.5B), 12B (unified encoder-free architecture), 26B-A4B (25.2B total / 3.8B active, Mixture-of-Experts), and 31B dense — all multimodal (text, image; audio natively on E2B/E4B/12B), with 128K context on the small models and 256K on the 12B and up, and multilingual support across 140+ languages (model card: https://ai.google.dev/gemma/docs/core/model_card_4). Independent tracking from Artificial Analysis puts the 31B reasoning variant at an Intelligence Index of 30, ranked #9 of 134 open-weight models it tracks in that size class, well above the median of 9 for similar-sized open models (https://artificialanalysis.ai/models/gemma-4-31b). The same source flags it as notably slow on raw output speed — 35.5 tokens/sec, ranked #66/134 — so it trades throughput for reasoning quality rather than winning on both. On the crowdsourced LMArena leaderboard, Google's own reporting places the 31B dense model at #3 among open models and the 26B MoE at #6 while activating under 4B parameters per token, based on blind human side-by-side ratings (https://huggingface.co/blog/gemma4).

There are three real paths to running Gemma, and they have different cost and privacy trade-offs. Self-hosting the open weights (via Hugging Face, Kaggle, Ollama, llama.cpp, or Unsloth) costs only your own compute — Gemma 4 weights are downloadable with no usage fee. Through Google's own Gemini API / AI Studio, Gemma 4 input, output and context-caching are all listed as free of charge with no paid tier offered at all — but the same pricing page states free-tier usage may be used to improve Google's products, which matters if you need a contractual no-training guarantee (checked directly on https://ai.google.dev/gemini-api/docs/pricing, page timestamped 2026-08-13). If you want a paid, presumably-no-training hosted endpoint instead, Vertex AI offers the Gemma 4 26B variant at $0.15 per million input tokens and $0.60 per million output tokens, with a $0.015 cache-hit rate (verified directly on https://cloud.google.com/vertex-ai/generative-ai/pricing via its Gemma pricing table) — notably, only the 26B size appears with hosted pricing there, not the other four sizes.

The reference google-deepmind/gemma GitHub repository (the inference library, not the weights) has 5,659 stars, 1,012 forks, and was last pushed to on 2026-08-10 — a week before this check — with 316 open issues, indicating active, ongoing maintenance rather than a one-off release dump (checked via GitHub API, https://github.com/google-deepmind/gemma). Individual model weights show heavy real usage on Hugging Face: the gemma-4-31B-it and gemma-4-26B-A4B-it repos each report roughly 9.7–10 million downloads and thousands of likes (checked via Hugging Face's model API). Day-one tooling support is broad — Kaggle, Hugging Face, Keras, Ollama, PyTorch, JAX, Google AI Studio, Vertex AI, Android, LM Studio and Unsloth are all listed as supported access points on the official DeepMind models page (https://deepmind.google/models/gemma).

In late October 2025, Google removed Gemma from AI Studio's web interface after Senator Marsha Blackburn said the model, asked a factual question about her, fabricated a detailed and false sexual-misconduct allegation with fake supporting article links (TechCrunch: https://techcrunch.com/2025/11/02/google-pulls-gemma-from-ai-studio-after-senator-blackburn-accuses-model-of-defamation/). Google's public response was that it had seen non-developers using AI Studio's Gemma option to ask it factual, current-events-style questions, and said plainly: 'We never intended this to be a consumer tool or model, or to be used this way.' The removal was partial — Gemma stayed available to developers through the API and other deployment channels — and was framed by Google as a UI/positioning fix, not a model recall. It's a useful data point independent of whether Gemma 4 or later fixed the underlying hallucination behavior: this model family is positioned and marketed for developers building products, not for direct end-user question-answering, and a buyer building a public-facing Q&A or chat surface on Gemma should weight that accordingly.

Google publishes a companion safety classifier, ShieldGemma 2, and offers content-safety filtering guidance in the model cards; the Gemma 4 model card documents its own ethics-and-safety evaluation approach and states standard prohibited-use categories (CSAM, dangerous instructions, explicit sexual content, hate speech) while explicitly telling integrators to add their own content-safety safeguards on top (https://ai.google.dev/gemma/docs/core/model_card_4). Independent academic red-teaming has found the smaller Gemma 3 variants (1B) substantially more vulnerable to prompt-injection and jailbreak techniques than the larger ones in the same family, per a 2026 workshop paper evaluating multiple open models (https://dl.acm.org/doi/10.1145/3803628.3807972) — a pattern that is common across small open models generally, not unique to Gemma, but worth knowing if you're deploying the smallest on-device sizes without your own guardrail layer.