On April 2, 2026, Google announced "Gemma 4", its flagship open-weight model. The release ships in four sizes including a 26B Mixture of Experts (MoE) and a 31B Dense model, and the headline numbers — release under the Apache 2.0 license and a high LMArena ranking — are getting a lot of attention. At the same time, even though it is billed as "open", there are practical constraints that operators need to keep in mind, and a calmer view that includes concerns about benchmark-first thinking is also warranted. This post lines up the Gemma 4 specs alongside the points worth debating.
Announcement: Four-Size Lineup and Benchmark Numbers
Gemma 4 is offered in four sizes: Effective 2B (E2B), Effective 4B (E4B), 26B Mixture of Experts (MoE), and 31B Dense. The 26B MoE activates only a portion of its total parameters at inference time (around 3.8B), prioritizing latency, while the 31B Dense model emphasizes raw quality and fine-tuning suitability.
Source: Google blog "Gemma 4: Byte for byte, the most capable open models"
On benchmarks, the 31B Dense model is reported at an estimated LMArena Text score of 1452, and the 26B MoE at around 1441, ranking 3rd and 6th respectively on Arena AI's open-model leaderboard. Numbers like 85.2 percent on MMLU Pro, 89.2 percent on AIME 2026, and 2150 on Codeforces also put it at a competitive level among open models.
Source: Google DeepMind "Gemma 4"
The license is Apache 2.0, with no monthly active user cap or additional policy acceptance for commercial use. This is a major shift from the bespoke licenses used for Gemma 1 through 3.
Bias in Evaluation Axes: The Pitfalls of Benchmark-First Thinking
Looking at the numbers alone, Gemma 4 is plenty strong, but there are several caveats. First, LMArena relies on human preference voting, and critics have pointed out for some time that it is heavily influenced by output style (verbosity, politeness, neat Markdown formatting). Models that optimize for "well-liked presentation" rather than answer quality itself tend to be overrated, so calling something "the strongest" based on top-of-leaderboard placement alone may be premature.
Second, problem sets like AIME and MMLU Pro carry a debated risk of contamination in pre-training data, and high scores on specific tasks do not guarantee comprehensive reliability in real-world use. In areas like coding or agent operation, where error detection and self-correction over long contexts matter, the gap between benchmark numbers and user experience tends to show up.
Source: Google blog "Gemma 4: Byte for byte, the most capable open models"
What "Open" Really Means and Operational Considerations
It has long been pointed out that the Gemma series continues to be tied to Google's "Gemma Prohibited Use Policy", and rather than treating it as "fully free because it is Apache 2.0", it is important to review both the license document and the policy before adoption.
On the performance side, the 26B MoE has small active parameters, but generalization — the stability of expert routing, behavior on Japanese tasks, and so on — is what really gets tested in the operations phase.
Gemma 4 is also positioned as the foundation for the next-generation Gemini Nano, and within Google's competitive strategy, it serves as a stepping stone for keeping on-device AI inside the company's own ecosystem, like Pixel. Distributing it for free as an open model while ultimately recouping the value as a differentiator for first-party products is a structure the wider AI community needs to look at carefully from an ecosystem-balance perspective.
Source: Google DeepMind "Gemma 4"
Summary
Gemma 4 is an open-weight model announced on April 2, 2026, and it draws attention with its four-size lineup, Apache 2.0 license, and high LMArena ranking. Once you factor in the limits of benchmark-centric evaluation, the attached policy, and the realities of operating costs, simplifications like "the strongest" or "fully free" may be premature, and anyone considering adoption should make sure to validate against their own tasks and read the license carefully.
