Gemma 4
Gemma 4 is an open-source large language model series released by Google, with a parameter scale of 31 billion (31B). Built on the Transformer architecture, it supports tasks including text understanding, generation, and multi-step reasoning. Model weights are made available to researchers and developers for local deployment and fine-tuning on downstream tasks. As an iteration of the Gemma series, it is designed to deliver strong general-purpose language processing capabilities while maintaining relatively low resource requirements.
Core Facts
Gemma 4's 26B MoE model uses a Mixture of Experts architecture, with 26 billion total parameters but only approximately 4 billion parameters active at any given moment.
Gemma 4 adopts the Apache 2.0 license, allowing completely unrestricted commercial use
Gemma 4的1B/2B模型量化后仅需4GB显存即可运行
Timeline (last 90 days)
All calls to Gemini, Gemma, Claude, and Nimble are Sentry spans, and background tasks each form their own transaction; every Gemini-to-Claude fallback also raises a warning issue; Sentry is configured with send_default_pii=False, so student photos and prompt text are not sent.
Gemma 4 (gemma-4-26b-a4b-it, going through the same Gemini API) handles high-frequency, low-risk tasks: grading and determining whether two answers are equivalent.
Somo is a math tutoring app that takes a photo upload of a handwritten solution process, uses AI to find the first line in the handwritten solution where an error occurs, circles it in red like a teacher, and explains it in English, Swahili, or Spanish.
The MCP server demo includes six inference backends: Amazon Bedrock, a SageMaker endpoint, vLLM on EC2 g6, an inference service on g5g, and Gemma 4 models on Inferentia2 (inf2) and Trainium (trn1).
In evaluations across 3 seeds, 3 models, and 4 benchmarks (involving Gemma 4, Qwen3.5, and Magistral), U-Lens is the best estimator, ahead of 5-round sampling, 10-round sampling, and supervised baselines.
On two-step reasoning questions, Claude Haiku 4.5 scored 6/30, GPT-5.4 mini scored 3/30, while Claude Sonnet 5.5, Gemini 3.8 Flash, and Gemma 4 all scored 30/30.
The free open-source model Gemma 4 answered all 60 answerable questions correctly and correctly labeled sources every time; gpt-oss-20b answered 59 correctly; the paid small models Claude Haiku 4.5 answered 36 correctly and GPT-5.4 mini answered 32 correctly.
Open-source models scored lower: Qwen3 235B scored 11/12, both sizes of Gemma 4 scored 9/12, DeepSeek R1 scored 7/12, and GLM-5 scored 5/12.
Both models expressed almost identical confidence when fabricating versus when answering correctly (Gemini 93.7% vs 95.5%, Gemma 96.3% vs 99.0%), indicating that confidence scores cannot distinguish correct answers from fabrications.
Gemma 4 is the cheapest vision model on Kaggle, costing only $0.34 to run all 239 items, while Gemini 3.7 Flash cost $2.70.
36 more timeline events
All Facts (20)
Gemma 4's 26B MoE model uses a Mixture of Experts architecture, with 26 billion total parameters but only approximately 4 billion parameters active at any given moment.
80%VerifiedGemma 4 adopts the Apache 2.0 license, allowing completely unrestricted commercial use
75%VerifiedGemma 4的1B/2B模型量化后仅需4GB显存即可运行
65%Unverified全部四种Gemma 4模型都支持文本和图片输入，Edge模型（E2B和E4B）还额外支持音频输入
90%UnverifiedGemma 4支持完全离线状态下的图像文字提取（OCR），采用视觉-语言多模态架构
90%UnverifiedGemma 4大模型上下文窗口可达256,000 Token
90%Unverified所有Gemma 4模型原生支持多模态（文本、图像、最多30秒音频输入），上下文窗口为128K Token，支持超过40种语言
85%UnverifiedGemma 4引入了多模态能力，Qwen3支持超长上下文，DeepSeek采用MoE（混合专家）架构
85%Unverified使用Gemma 4系列模型验证了模型权重与Harness状态的联合学习效果，发现单独的监督微调和离线强化学习预热都不能让模型取得里程碑进展，只有联合学习循环才能产生持续提升
85%UnverifiedGemma 4的1B/2B模型自带语音识别功能
80%UnverifiedGoogle在Gemma 4中实际应用了MTP技术
80%UnverifiedGemma 4使用滑动窗口注意力（SWA）架构，该架构在处理数千行混淆JavaScript时导致模型遗忘早先读入的关键上下文
80%Partially VerifiedEmbeddingGemma 2 is a multimodal embedding model built by Google on the Gemma 4 architecture, capable of encoding text, images, audio, and video into a shared 768-dimensional vector space.
75%Unverified使用Gemma 4 2B版本本地运行Codex时，处理速度大约每秒30到40个Token
75%Partially VerifiedGoogle released Embedding Gemma 2, its first native multimodal open embedding model targeting on-device use, with only 740 million parameters, open weights under the Apache 2.0 license, capable of understanding text, code, images, audio, and video within a single shared space.
75%UnverifiedAll calls to Gemini, Gemma, Claude, and Nimble are Sentry spans, and background tasks each form their own transaction; every Gemini-to-Claude fallback also raises a warning issue; Sentry is configured with send_default_pii=False, so student photos and prompt text are not sent.
50%UnverifiedSomo is a math tutoring app that takes a photo upload of a handwritten solution process, uses AI to find the first line in the handwritten solution where an error occurs, circles it in red like a teacher, and explains it in English, Swahili, or Spanish.
50%UnverifiedGemma 4 (gemma-4-26b-a4b-it, going through the same Gemini API) handles high-frequency, low-risk tasks: grading and determining whether two answers are equivalent.
50%UnverifiedThe MCP server demo includes six inference backends: Amazon Bedrock, a SageMaker endpoint, vLLM on EC2 g6, an inference service on g5g, and Gemma 4 models on Inferentia2 (inf2) and Trainium (trn1).
50%UnverifiedIn evaluations across 3 seeds, 3 models, and 4 benchmarks (involving Gemma 4, Qwen3.5, and Magistral), U-Lens is the best estimator, ahead of 5-round sampling, 10-round sampling, and supervised baselines.
50%