Google's Gemma 4: An American Challenge to China

This week, Google DeepMind introduced Gemma 4, an open-weight model that marks a significant advancement for the United States in the field of artificial intelligence, a sector recently dominated by China. For several months, Chinese models, particularly those based on Mixture-of-Experts architectures, have captured attention due to their increasing size and complexity. While Gemma 4 does not overturn this dominance, it reintroduces a robust family of models under the Apache 2.0 license, allowing execution on local hardware or in stricter enterprise environments.

However, the market share emphasizing self-hosting is declining. Anthropic recently announced that its annual revenue has reached $30 billion, up from about $9 billion at the end of 2025 and around $1 billion in December 2024. This represents a 30-fold increase in just 16 months. More and more clients are becoming comfortable with using LLM APIs or enterprise-level agents and chatbots, compared to six months ago. The security and privacy policies of major AI labs have become clearer, which has helped lower the barrier for risk-averse organizations.

Google has launched four variants of Gemma 4: the E2B and E4B models for lightweight applications, the flagship dense model of 31B, and a 26B MoE A4B model designed for high-throughput reasoning. Gemma has now surpassed 400 million downloads and over 100,000 community variants. This generation is built on the research from Gemini 3 and, for the first time, is offered under the Apache 2.0 license.

On Google’s benchmarks, the two larger models show serious performance. The 31B model scores 1,452 on Arena AI text, 84.3% on GPQA Diamond, 89.2% on AIME 2026, 80.0% on LiveCodeBench v6, 76.9% on MMMU Pro, and 86.4% on Tau2-bench retail, compared to just 6.6% for Gemma 3 27B on the same test. The 26B A4B model closely follows with a score of 1,441 on Arena AI text, 82.3% on GPQA Diamond, 88.3% on AIME 2026, and 77.1% on LiveCodeBench. Google also reports scores of 19.5% and 8.7% on Humanity’s Last Exam without tools for the 31B and 26B models, respectively, rising to 26.5% and 17.2% with search. These results position these open models as genuinely competitive.

The architecture of Gemma 4 is conservative, which adds to its appeal. It uses a hybrid sliding window combined with global attention, as well as Proportional RoPE for long context, with a local window of 512 tokens on the lightweight models and 1,024 tokens on the larger models. The 31B model has 30.7 billion effective parameters, while the 26B A4B has a total of 25.2 billion, but only 3.8 billion active per token (8 out of 128 experts plus one shared). The improvement in capabilities seems to be more due to reinforcement learning, training recipes, and data rather than architectural reinvention.

On the engineering side, Gemma 4 supports configurable thinking mode, native system role prompting, native function calling with dedicated tokens, and text-image input across the family, as well as video and audio on the smaller models. The prompting documentation is exceptionally concrete, with a clearly defined tool lifecycle, direct advice on removing thought traces from multi-turn history, and a recommendation to summarize reasoning in context for long-term agents rather than replaying raw tokens. Google also explicitly warns developers to validate function names and arguments before execution.

The smaller models target phones, Raspberry Pi, and Jetson Nano, while the 26B and 31B models are suited for consumer GPUs and workstations. Both larger models can run on a single H100. An important caveat: despite having only 3.8 billion active parameters, the 26B MoE still requires loading the full model into memory. The MoE does not offer a free deployment. Ecosystem support is comprehensive: available from day one on Hugging Face, Ollama, Kaggle, LM Studio, vLLM, llama.cpp, MLX, NVIDIA NIM, Vertex AI, and Google AI Edge. On Android, Gemma 4 serves as the foundation for Gemini Nano 4, offering up to four times the performance and 60% less battery consumption.

The independent analysis from Artificial Analysis is nuanced. On its Intelligence Index, the 31B model scores 39, trailing behind Qwen 3.5 27B at 42 by just 3 points while using about 2.5 times fewer output tokens to complete the benchmark suite (39M vs. 98M). The main weakness of the 31B compared to Qwen is its agentic performance, not general reasoning. In non-agentic evaluations, it is on par: SciCode 43 vs. 40, TerminalBench Hard 36 vs. 33, GPQA Diamond 86 vs. 86, IFBench 76 vs. 76, Humanity’s Last Exam 23 vs. 22. The 26B A4B presents a less flattering story, clearly trailing behind Qwen 3.5 35B A3B on agentic work (Agentic Index 32 vs. 44). In summary: the 31B is the star, the 26B A4B is useful but not magical, and the smaller models outperform their weight.

Gemma 4 is important because it changes the dynamics of the open-weight market, not because it takes the crown. Last year's dominance of Chinese labs produced brilliant models, but many are trillion-parameter MoE systems that are difficult to self-host, costly to run properly, and, for some Western companies, uncomfortable from a compliance perspective. Gemma 4 offers these organizations a credible alternative: American-made, under Apache 2.0, practical to deploy on a single GPU. For regulated sectors, isolated environments, edge devices, and teams that need control over data retention and customization, it is a real option, not a toy.

At the same time, Anthropic's annual revenue of $30 billion is solid evidence that the broader market is moving towards hosted APIs and enterprise-level products rather than self-hosting. I believe this reduces the role of open weights, but it also clarifies it. Open models no longer need to serve everyone. They must own the use cases where locality, inspectability, and tuning flexibility matter more than the frontier of capabilities.

It is also worth noting that the AI engineering space has continued to move away from fine-tuning. Most production teams rely entirely on prompting, retrieval, and context engineering, and leading closed models are generally not available for fine-tuning at the weight level anyway. The bar for fine-tuning a small open model to surpass the out-of-the-box capabilities of a leading model with good tools and context is extremely high. But Gemma 4 is important here precisely because it maintains a credible path for customization for teams that genuinely need it, at a much higher capability level than previous American open-weight options.

My broader view: the likely future is not open versus closed. It’s hybrid. APIs or leading agents where they are clearly best, open weights where locality, privacy, predictable cost, or customization outweigh capabilities. Teams building for both sides of this trade-off will do well.