Google Jumps Back Into the Open Source AI Race With Gemma 4

Google's unfastened AI ambitions got a batch much superior today. The institution released Gemma 4, a household of 4 open-weight models built connected the aforesaid probe arsenic Gemini 3, and licensed nether Apache 2.0—a important departure from the much restrictive presumption connected erstwhile Gemma versions.

Developers person downloaded past Gemma generations implicit 400 cardinal times, spawning much than 100,000 assemblage variants. This merchandise is the astir ambitious 1 yet.

For the past year, the open-source AI leaderboard has been mostly a Chinese affair. DeepSeek, Minimax, GLM and Qwen person dominated the apical spots, leaving American alternatives scrambling for relevance. As Decrypt reported past year, Chinese unfastened models went from hardly 1.2% of planetary open-model usage successful precocious 2024 to astir 30% by the extremity of 2025, with Alibaba's Qwen adjacent overtaking Meta's Llama arsenic the most-used self-hosted exemplary worldwide.

Meta's Llama utilized to beryllium the default prime for developers who wanted a capable, locally runnable model. That estimation has eroded—Llama's Meta-controlled licence raised questions astir its existent open-source status, and its show slipped down the Chinese competition. The Allen Institute's OLMo household tried to capable the spread but failed to summation meaningful traction. OpenAI released its gpt-oss models successful August 2025, which gave the ecosystem a enactment of caller air, but they were ne'er designed to beryllium frontier competitors.

And yesterday, a 30-person U.S. startup called Arcee AI released Trinity, a 400 cardinal parameter unfastened exemplary that made a compelling lawsuit that the American country wasn't wholly dead. Gemma 4 follows that momentum, this clip with the afloat value of Google DeepMind down it, turning it into arguably the champion American exemplary successful the open-source AI scene.

The exemplary is "built from the aforesaid world-class probe and exertion arsenic Gemini 3," Google said successful its announcement. Gemma 4 ships successful 4 sizes: Effective 2B and 4B for phones and borderline devices, a 26B Mixture of Experts exemplary focused connected speed, and a 31B Dense exemplary optimized for earthy quality.

The 31B Dense presently ranks 3rd among each unfastened models connected Arena AI's substance leaderboard. The 26B MoE sits sixth. Google claims some outcompete models 20 times their size—a assertion that holds up, astatine slightest against the Arena AI numbers, wherever Chinese models inactive clasp the apical 2 spots.

We tested Gemma 4. It's capable, with immoderate caveats. The exemplary applies reasoning adjacent to tasks that don't necessitate it, which tin marque responses consciousness over-engineered for elemental prompts. Creative penning is decent—serviceable, not inspired—and apt improves with much circumstantial guidance and punctual engineering.

Where it delivered astir intelligibly was code. Asked to make a game, the output wasn't peculiarly flashy oregon elaborate, but it ran without errors connected the archetypal try. Not atrocious for a 41 cardinal parameter model. That zero-shot reliability is arguably much invaluable than a prettier effect that needs debugging.

You tin effort the (basic, yet functional) crippled here.

The 4 variants screen the afloat hardware spectrum. The E2B and E4B models are built for Android phones, Raspberry Pi, and borderline devices, moving wholly offline with near-zero latency, autochthonal audio input, and a 128K discourse window. The 26B and 31B models people workstations and unreality deployments, extending discourse to 256K and adding autochthonal function-calling and structured JSON output for gathering autonomous agents. All 4 models process images and video natively. The larger models' full-precision weights acceptable connected a azygous 80GB NVIDIA H100 GPU; quantized versions tally connected user hardware.

The Apache 2.0 licence is the different headline. Google's erstwhile Gemma releases utilized a customized licence that created ineligible ambiguity for commercialized products. Apache 2.0 removes that friction entirely—developers tin modify, redistribute, and commercialize without worrying astir Google changing the presumption later. Hugging Face co-founder Clement Delangue praised it, saying that "Local AI is having its moment," and it is the aboriginal of the AI industry. Google DeepMind CEO Demis Hassabis went further, calling Gemma 4 "the champion unfastened models successful the satellite for their respective sizes."

That's a beardown claim. Proprietary systems from Anthropic, OpenAI, and Google's ain Gemini inactive pb connected the hardest benchmarks. But for open-weight models you tin tally locally, modify freely, and deploy connected your ain infrastructure? The contention conscionable got importantly thinner. You tin effort Gemma 4 present successful Google AI Studio (31B and 26B) oregon Google AI Edge Gallery (E2B and E4B). Model weights are besides disposable connected Hugging Face, Kaggle, and Ollama.