Gemma (Google open-weight models)

Gemma is Google DeepMind's family of open-weight models, positioned as Google's open-weight offering alongside Meta's Llama and Mistral's open-weight releases. It is distinct from Google's closed-source Gemini family, with a different architecture, different weights, and a different license: Gemma weights are released openly, while Gemini is closed. Google has described Gemma as powering enterprise fine-tuning and on-device deployments. With Gemma 4 in April 2026 the family moved from a bespoke Google licence to Apache 2.0, a change of interest to the open-weight policy debate independent of the models' capabilities.

Generations

The family has advanced through several generations since its first release:

• Gemma 1 (February 2024): 2B and 7B variants.
• Gemma 2 (2024): 9B and 27B variants.
• Gemma 3 (March 2025): additional variants and multimodal extensions; Ars Technica described it as optimized to run on a single GPU.
• Gemma 4 (April 2, 2026): four sizes, all released under Apache 2.0 (Source: arstechnica.com).
• DiffusionGemma (June 10, 2026): an experimental open 26B mixture-of-experts model (3.8B active parameters) built on a Gemma 4 backbone that uses text diffusion to generate blocks of text in parallel rather than strictly left-to-right, reported to deliver up to roughly 4× faster GPU inference (Source: blog.google; siliconangle.com). Its parameter count and active-parameter count match the Gemma 4 26B MoE configuration.

Conflicting dates for Gemma 4 in secondary reporting

Two newsletter items date the Gemma 4 release differently from Google and from Ars Technica, and the discrepancy is recorded rather than resolved in their favour. A newsletter roundup described a "Gemma 4 12B" released around June 2, 2026 as an encoder-free unified multimodal open model under Apache 2.0, fitting in 16GB of memory, running agentic reasoning and native audio on a laptop, and approaching the performance of Google's larger 26B mixture-of-experts model at less than half the memory footprint, in the same early-June window as NVIDIA's open Nemotron 3 Ultra (550B) and Microsoft's MAI slate (Source: nlp.elvissaravia.com). A second roundup, for the week ending July 11, 2026, reported that Google open-sourced Gemma 4 that week, listing it alongside Tencent's open-weight Hy3 release (Source: nlp.elvissaravia.com).

Google's own announcement and Ars Technica's contemporaneous report both place the Gemma 4 launch on April 2, 2026, and neither lists a 12B variant among the four sizes (Source: arstechnica.com; opensource.googleblog.com). The April date is treated here as established. The 12B model is corroborated by Google's own documents: the Gemma 4 model card, updated July 30, 2026, lists five sizes — E2B, E4B, 12B, 26B A4B and 31B — and describes the "12B Unified" model as 11.95 billion parameters with 48 layers (Gemma 4 Model Card (Google DeepMind, 2026)); the Gemma 4 technical report of July 24, 2026 lists dense models of 2.3B, 4.5B, 12B and 31B parameters and introduces "a unified encoder-free architecture for the 12B model" (Gemma 4 Technical Report (Google DeepMind, July 2026)). The 12B therefore appears to be a later addition to the April lineup rather than a newsletter error.

Gemma 4

Gemma 4 is built on the same underlying technology as Google's closed Gemini 3 models and offers, per Google, improved reasoning, mathematics and instruction-following relative to Gemma 3. It adds native function calling, structured JSON output and native instructions for common tools and APIs, and Google positions it for offline code generation and for visual tasks such as OCR and chart understanding (Source: arstechnica.com).

The four sizes split by deployment target. The 26B Mixture-of-Experts and 31B dense variants are designed to run unquantized in bfloat16 on a single 80GB Nvidia H100, and to fit consumer GPUs when quantized; the 26B MoE activates 3.8 billion of its 26 billion parameters at inference, which Google presents as yielding higher throughput than similarly sized dense models, while the 31B dense variant is positioned for quality and fine-tuning. The Effective 2B and Effective 4B variants target mobile devices, maintaining effective 2-billion and 4-billion parameter memory footprints during inference; Google says its Pixel team worked with Qualcomm and MediaTek to optimize them for smartphones, Raspberry Pi and Jetson Nano, and claims lower memory and battery use than Gemma 3 alongside "near-zero latency." Both edge variants carry native speech recognition, a capability the Gemma 3 family also had (Source: arstechnica.com).

Google stated at launch that Gemma 31B would debut at number three on the Arena leaderboard of open models, behind GLM-5 and Kimi 2.5, and noted that the largest Gemma 4 variant is a fraction of the size of either (Source: arstechnica.com). The ranking claim is Google's own characterization of a third-party leaderboard at a point in time and is a fast-decay figure.

Nathan Lambert wrote in September 2026 that the Gemma 4 models have been "some of the only ones ever" to show adoption numbers similar to Qwen's most popular small models (The current balance of power in open models (Nathan Lambert, September 2026)).

Licensing

Gemma 1 through 3 were released under the Gemma Terms of Use, a bespoke Google licence. Ars Technica characterized developer objections to it in specific terms: a prohibited-use policy Google could update unilaterally, a requirement that developers enforce Google's rules across all Gemma-derived projects, and language that could be read to extend the licence to other models trained on synthetic data produced by Gemma (Source: arstechnica.com). Google's developer documentation retains the Gemma Terms of Use for earlier generations, last modified April 1, 2026, and directs Gemma 4 users to the Apache 2.0 licence (Source: ai.google.dev).

Apache 2.0 removes the use restrictions and the unilateral-amendment mechanism, and permits commercial use, fine-tuning, modification and redistribution of derivatives without a bespoke review (Source: arstechnica.com). The licence covers the weights; it does not by itself make training data or training code available, so the release remains open-weight rather than open-source in the sense the Open Source Initiative applies to AI systems.

Relation to Gemini Nano

Google's on-device Gemini Nano models are derived from Gemma. Google confirmed to Ars Technica that Gemini Nano 3, running on Pixel phones, is based on Gemma 3n, and that a next-generation Gemini Nano 4 will ship 2B and 4B variants based on Gemma 4 E2B and E4B. Developers were invited to prototype agentic workflows against E2B and E4B in the AI Core Developer Preview with forward compatibility to Nano 4 (Source: arstechnica.com). This makes the Gemma line the upstream of the local models handling scam detection, note summarization and call summaries on Android devices without sending data to the cloud.

Evaluations

Gemma 1.1 2b IT, the instruction-tuned variant, was rated the least politically biased conversational LLM in an analysis by David Rozado published in January 2025 (Manhattan Institute — Measuring Political Preferences in AI Systems (Rozado)).

No system card or published safety case for the Gemma family has been located, and the safety_case field is recorded as none-published pending one.

Relationships

• related: The current balance of power in open models (Nathan Lambert, September 2026) — Gemma 4 adoption (Lambert, September 2026)
• developer: Google DeepMind
• sibling-family: Gemini 3 / Gemini 3 Pro, Gemini 2.5 Pro (closed-source Gemini family)
• instance-of: Open-Weight Frontier Models — an open-weight release from a frontier lab, and the licence change is a datapoint in that debate
• related: Open-Source AI / Open-Weight Models, Llama 4 (Meta), Manhattan Institute — Measuring Political Preferences in AI Systems (Rozado), GLM-5.2, NVIDIA Nemotron.
• documented-by: Gemma 3 Technical Report (Google DeepMind, March 2025), Gemma 4 Technical Report (Google DeepMind, July 2026), Gemma 4 Model Card (Google DeepMind, 2026)