Gemma
Gemma is a series of open-source large language models released by Google, built on the same research and technology underlying Gemini. The series is available in multiple parameter sizes and supports natural language processing tasks including text generation, question answering, and code assistance. Designed with lightweight efficiency as a goal, Gemma supports local deployment on consumer-grade hardware, and developers can obtain model weights at no cost for fine-tuning and further development.
Core Facts
Gemma is a series of open-source models developed by Google DeepMind based on Gemini model technology.
开放权重模型指参数权重文件向公众开放下载，可本地部署微调，但训练数据和完整源码不一定公开，区别于完全开源
DiffusionGemma is an experimental model that fuses the Google Gemma architecture with the diffusion generation paradigm.
Timeline (last 90 days)
Gemma's draft model uses multi-token prediction (MTP) speculative decoding, sharing the base model's KV cache, achieving an acceptance rate 40 to 60% higher than most draft models, 40% lower decoding latency, and 30 to 40% lower energy consumption.
Some time has passed since the last update of the Gemma series.
Past Gemma releases have consistently followed a cycle of "first Gemini, then Gemma."
Gemma is an open-source model series released by Google; 26B means roughly 26 billion parameters, and even after quantization it still requires over ten to more than twenty GB of memory, so the first barrier to local inference is often not compute power but memory capacity.
The author runs the 26B version of Gemma 4 (the MLX version) on a MacBook Pro with 36GB of memory for local AI inference.
In practical tasks, cloud models (such as the Claude desktop version) perform more reliably in work such as fixing broken automations and generating dashboards, while local models require more guidance.
Common local small-model options include Meta's small Llama variants, Microsoft's Phi series, Google's small Gemma models, and loading locally via runtimes such as llama.cpp and Ollama.
When running Gemma with Ollama, Cuadre's only network connection is localhost:11434, and the code does not accept any other address, ensuring receipt data does not leave the local machine.
Cuadre's core design principle is "the model transcribes, the code calculates": Gemma is only responsible for copying the printed content on the receipt, and all arithmetic, validation, and summarization are done by ordinary Python code.
Cuadre is an open-source project that uses a locally running Gemma model to automatically extract information from photos of Panamanian receipts and generate tax tables (CSV, summary Markdown, and a to-review checklist).
40 more timeline events
All Facts (20)
Gemma is a series of open-source models developed by Google DeepMind based on Gemini model technology.
90%Verified开放权重模型指参数权重文件向公众开放下载，可本地部署微调，但训练数据和完整源码不一定公开，区别于完全开源
80%VerifiedDiffusionGemma is an experimental model that fuses the Google Gemma architecture with the diffusion generation paradigm.
80%VerifiedOllama是一款开源的本地大模型运行框架，支持在个人电脑上一键部署各种开源大语言模型
80%VerifiedGemma is a series of open-weight language models launched by Google in 2024; its name derives from the Latin "gemma" (gem) and shares the same origin as Gemini but has a different positioning: Gemini is Google's flagship closed-source model for cloud deployment, while Gemma focuses on running in environments with limited compute.
75%VerifiedThe model weights alone of a 175-billion-parameter model (in BF16 precision) require 350 GB of storage, and the Adam optimizer must maintain two sets of states—first-order moments and second-order moments—totaling more than 1 TB.
75%VerifiedOpen weight means the model parameters can be downloaded and executed locally, but the training code, training data, and terms of use are not necessarily open; Gemma uses a custom Gemma Terms of Use.
65%VerifiedOllama is a local large model runtime tool that listens on port 11434, requires no authentication by default, and provides REST APIs for model management and inference including /api/pull, /api/create, and /api/chat.
65%VerifiedGemma is a lightweight open-source model series introduced by Google, emphasizing high efficiency under limited computing resources.
65%VerifiedThe current open-source large model landscape has formed a multipolar competitive structure, including Meta's Llama, Alibaba's Qwen, Mistral AI, and DeepSeek.
65%VerifiedGemma is a family of lightweight open-source language models released by Google in 2024, with parameter scales ranging from 2B to 7B.
65%VerifiedMeta's LLaMA series, Google's Gemma, Mistral AI's Mistral/Mixtral, and Alibaba's Qwen series are all supported in Transformers.
65%VerifiedThe Gemma series of models uses the SigLIP vision encoder to process image inputs.
65%VerifiedGoogle DeepMind has released Embedding Gemma 2, an open multimodal embedding model with 740M parameters that maps text, images, video, and audio into a shared vector space.
70%VerifiedGemma is a family of open-source lightweight language models released by Google DeepMind in early 2024, refined through knowledge distillation technology.
65%Unverified可用于本地部署的模型从轻量级的Phi-3、Gemma到重量级的Llama 3、Qwen2.5都可以根据硬件条件自由切换
90%UnverifiedMeta 的 LLaMA 系列、Mistral AI 的 Mistral/Mixtral、Google 的 Gemma 等开源模型都选择通过 Hugging Face 平台和 Transformers 格式进行发布
90%Unverified微软Phi-3最小版本仅3.8B参数，Google Gemma有2B和7B两个规格
90%Unverified小语言模型（SLM）通常指参数量在数亿到数十亿级别的模型，代表包括Microsoft的Phi系列和Google的Gemma系列
90%Unverified微软的Phi系列小模型和Google的Gemma都大量使用了合成数据进行训练
85%