In October, a prototype satellite from Google’s Project Suncatcher sent back a haiku. Per Google’s October recap, it was “the first-ever AI response generated by a data center-class Google TPU in space,” and the model that wrote it was Gemma 4. A stunt, yes. It also shows how far an open model can travel once nobody has to clear a license before moving it somewhere new.
When Google Gemma 4 shipped on April 2, this site called the Apache 2.0 license the real story. Seven months on, that call holds up. Several of the reasons we gave for it did not. The original version of this post got the old Gemma license wrong, overstated where Gemma 4 sat on the leaderboards, and linked to a Hugging Face repository that never existed. This is the corrected account, with what has happened since.
What we got wrong in April
The old license. We wrote that the Gemma Terms of Use carried “monthly active user thresholds.” They did not. The user cap belongs to Meta: the Llama 4 Community License requires any licensee with more than 700 million monthly active users in the preceding calendar month to “request a license from Meta, which Meta may grant to you in its sole discretion.” The Gemma Terms of Use have no user count at all. What they do have is harder for a legal team to live with. Section 3.2 incorporates Google’s Prohibited Use Policy by reference, section 3.1 makes anyone who redistributes a Gemma model pass those restrictions on to their own users “as an enforceable provision,” and section 3.2 also says that “Google reserves the right to restrict (remotely or otherwise) usage of any of the Gemma Services that Google reasonably believes are in violation of this Agreement.”
That last sentence was the friction. A cap of 700 million users never touched most enterprises. A vendor’s reserved right to restrict use, judged by what the vendor “reasonably believes,” touches everyone who embeds the model in a product. Those terms still exist (last modified April 1, 2026) and still govern Gemma 1, 2, 3, 3n, EmbeddingGemma, PaliGemma, ShieldGemma and the rest of the older family. The page now says only: “For Gemma 4 terms, see the Gemma 4 license,” and that link resolves to the plain Apache 2.0 text, with no use policy and no extra clauses.
The leaderboard position. We implied Gemma 4 31B was the best open model in its size class. On the independent Artificial Analysis Intelligence Index, it scored 39 against 42 for Qwen3.5 27B (Reasoning). Its Arena standing (3rd among open models for the 31B, 6th for the 26B MoE, per Wikipedia’s summary of the launch rankings) is a human preference vote, not a capability ranking, and we should have said so.
Coding. We said Qwen led Gemma on code. That is roughly true at the family level, but Artificial Analysis found the 31B ahead of Qwen3.5 27B on SciCode (43% vs 40%) and TerminalBench Hard (36% vs 33%), and attributed the overall gap “primarily due to weaker agentic performance.” If you are choosing between them for an agent that calls tools for twenty steps, that distinction matters more than any single coding score.
Smaller errors. We said the E2B “fits in under 1.5 GB quantized.” Artificial Analysis says the 4-bit weights fit “in under 3GB of RAM.” We linked readers to a gemma-4-27b-it repository that returns an error; the real checkpoints are gemma-4-26B-A4B-it (25.2B total, 3.8B active) and gemma-4-31B-it (30.7B dense). We also claimed a single consumer GPU would cut per-query cost to zero, which ignores power, operations staff and the hours the card sits idle. Those lines are gone.
The number April missed: tokens per answer
The most useful figure from the launch was not a benchmark score. Artificial Analysis reported that Gemma 4 31B used 39 million output tokens to run its index against 98 million for Qwen3.5 27B, roughly 2.5 times fewer. In non-reasoning mode the 31B scored 32 while spending just 7.1 million tokens, according to the firm’s sub-32B comparison.
On your own hardware, output tokens are GPU seconds. A model that scores three points lower but finishes in less than half the tokens serves more requests per card at a lower latency, and for classification, extraction and summarization work the three points rarely show up in the result. For reasoning-heavy agent loops, they can. That trade is the actual decision, and a single leaderboard number hides it.
The same comparison flags the weakness that matters most for enterprise retrieval and Q&A: factual recall. On AA-Omniscience, which penalizes confident wrong answers, Gemma 4 31B scored -45 and Qwen3.5 27B -42, against -10 for GPT-5 at medium and low effort. Small open models guess. Ground them in your own documents or expect to catch the misses downstream; our guide to building your own evals covers how to measure that before it reaches a user.
What shipped after April
Google kept filling in the family rather than replacing it.
- Gemma 4 12B, June 3. An 11.95-billion-parameter model under Apache 2.0 that VentureBeat reported runs on a laptop with 16GB of VRAM or unified memory. Google’s announcement describes it as encoder-free: the vision encoder replaced by a lightweight embedding module and the audio encoder removed entirely. On the Gemma 4 model card table it sits between E4B and the 26B MoE (77.2% MMLU Pro, 72.0% LiveCodeBench v6).
- A technical report in July. The model cards now cite a Gemma 4 technical report on arXiv dated July 2, which replaces launch-day marketing with documented architecture details.
- One billion downloads, August 20. Google DeepMind said the Gemma family had passed one billion cumulative downloads and more than 100,000 developer variants. TNW’s coverage adds the context the headline lacks: Google did not break the figure down by model or platform, Google’s Gus Martins noted on Bluesky that it “doesn’t count Android and Chrome integrations,” and Alibaba had claimed 3 billion Qwen downloads five days earlier. Hugging Face’s own count put Qwen at 2,045 million downloads this year on its hub alone. Gemma is a major open family. It is not the most downloaded one.
- EmbeddingGemma 2, October. A 740 million parameter embedding model built on the Gemma 4 architecture that maps text, code, images, video and audio into one vector space on device, with a modular design Google says cuts local vector storage “by up to 6x” (October recap). For anyone building private search, this may matter more than any chat model in the family.
What did not ship is a larger Gemma. The open frontier moved to trillion-parameter mixture-of-experts models, Mistral Large 4 and Kimi among them, while Gemma stayed in the lane where one GPU or one laptop does the job. That looks deliberate. Google’s frontier work goes into Gemini; Gemma is the part it gives away.
Apache 2.0 looks better in November than it did in April
The licensing case gets stronger the more open-weight releases you put next to it. Here is how the terms compare for the models an enterprise team is most likely to shortlist today:
| Model family | License | What it asks of you |
|---|---|---|
| Gemma 4 (all sizes) | Apache 2.0 | Keep the notice and license text. Nothing else. |
| Gemma 3 and earlier | Gemma Terms of Use | Follow the Prohibited Use Policy, pass it to downstream users, accept Google’s reserved right to restrict use |
| Llama 4 | Llama 4 Community License | Separate license above 700M MAU, an Acceptable Use Policy, “Built with Llama” display |
| Mistral Large 4 | Mistral Ecosystem Open-Weight License | Separate commercial license if you run a Model as a Service business |
The table is our summary of each license’s text, not a legal opinion. The Mistral row is the one worth dwelling on. As we covered when Mistral Large 4’s weights landed on October 29, its license requires a separate agreement if you give third parties inference access with “meaningful control” over inputs or parameters, with no revenue or user threshold. A carrier or managed service provider building a sovereign AI endpoint for business customers is precisely the case it describes.
Gemma 4 has no such clause. That is the practical meaning of Apache 2.0 for a service provider: you can host the model for customers, fine-tune it per customer and ship modified weights inside an appliance, and the only document your lawyers have to read is one they already know.
I have sat through the version of this review that goes badly. On infrastructure work at a large telecom, the slow part of putting an open model into a customer-facing service was never the GPU order. It was the weeks of back and forth over what a vendor’s acceptable use language meant once our customers, not our staff, were the ones typing prompts, and who would be liable if the vendor decided a use case crossed a line. A license that hands you the weights and asks only for an attribution file removes that whole workstream. That is worth more than three points on an index.
Where Gemma 4 fits on a shortlist now
For a team choosing an open model this quarter, the family sorts by lane, not by rank:
- Laptop or edge, offline: E2B, E4B or the 12B. Native audio input at these sizes is still rare, and the 12B’s 16GB footprint covers a standard enterprise laptop.
- One server GPU, high volume, short answers: the 26B MoE or the 31B in non-reasoning mode, where the token efficiency pays for itself. Artificial Analysis notes both the 31B and Qwen3.5 27B fit on a single 80GB H100 in BF16.
- Long agent loops with many tool calls: test Qwen3.5 27B alongside Gemma; the agentic gap is real. If the job needs frontier-level reasoning, neither belongs on the list, and the rent-versus-run math we did for Muse Spark Open applies.
- Private retrieval and search: EmbeddingGemma 2 for the index, a Gemma 4 model for the answer, and an eval harness that scores factual misses separately.
- Managed but inside a jurisdiction: Google’s Cloud launch post lists Gemma 4 across its Sovereign Cloud range, from public cloud with Data Boundary to Google Distributed Cloud “for air-gapped and on-premises deployments.”
Before any of these goes to production, read our notes on where open models fail when someone else hosts them and on what self-hosting actually costs at scale. An Apache license settles who may run the model. It does not settle who keeps it patched, monitored and isolated, and on your own hardware that job is yours.
The April version of this post ended by telling readers the gap between open and closed models had become small enough that license terms mattered more than benchmark deltas. The second half of that sentence survived. The first half needs a qualifier: the gap is small for the work small models do. For that work, Gemma 4 remains the open family a legal team approves fastest, and in November that is still the reason to put it on the list.