[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-gemma-4-qat-e2b-1gb-ram-quantization":3,"topics-all":36,"news-related-2e13fe83-ae62-422d-8ad9-526c97fcfd40":55},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"2e13fe83-ae62-422d-8ad9-526c97fcfd40","Gemma 4 QAT 让 E2B 跑进 1GB 内存：Google 把「训练时量化」做成消费级 AI 的最后一块拼图","把一个大模型塞进手机内存，过去两年的主流做法是训练完再压——PTQ。但当权重压到 4-bit 以下，PTQ 几乎都会带来明显的质量损失。Google 在 Gemma 4 上给出的答案是 QAT：把量化算子写进训练循环，让模型在权重更新的每一轮都预演自己被压扁后的样子，等到出 checkpoint 时就天然适合低位推理。为了兼顾 PC 端与移动端，Google 把 QAT 分成了两套配方。E2B\u002FE4B 这一档是为手机重做一套格式：静态激活省掉每次推理的动态 scale 计算；通道级量化对齐 NPU 的 memory layout；token 生成那部分 2-bit 极致压缩，核心推理层保留更高精度；embedding 与 KV cache 也单独压缩，让长上下文对话不会爆内存。E4B 以上的端侧\u002F服务器档则走标准 Q4_0，确保 llama.cpp、Ollama、LM Studio、vLLM、MLX 一条龙可用。真正的关键指标是 Gemma 4 E2B（文本-only，去掉 PLE）只需要 \u003C 1GB 内存。这基本意味着任何一台 8GB RAM 的手机、一块入门级 Apple Silicon 都能跑。Google 同时给了 MTP（Multi-Token Prediction）版本的 QAT 检查点，让推理加速和权重压缩叠加生效。Gemma 4 QAT 的价值不在某个单一数字，而在于训练—压缩—分发—部署被第一次系统性地打通。下一步消费级 AI 的瓶颈不再是模型能力，而是能不能在不联网、不烧电池的前提下，把一个真能用的助手塞进 1GB 的 RAM。Google 把这条链拆开了，而开源生态（llama.cpp、Unsloth、LiteRT-LM）随即跟进，意味着 4-bit 以下的本地 LLM 不再是 geek 玩具，而是真的要进入应用清单的选项。","https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Ftechnology\u002Fdevelopers-tools\u002Fquantization-aware-training-gemma-4\u002F","4d11edad-2df6-45f6-b71f-70f65de7f7fd",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"2d9c2fb0-2be5-4ad1-aedb-e9747addf355","compression",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":18,"name":19,"slug":19,"description":13,"color":13},"8cf7490f-2449-4ba7-be19-61befa0d92b4","google",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b49648f9-963e-4082-8684-3d085b7358fe","quantization",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"f03cfe65-64a4-4dc6-b5a5-ba60e704b5cf","en","Gemma 4 QAT squeezes E2B into 1GB of memory","Google AI Blog released Gemma 4 QAT (Quantization-Aware Training), a method that produces highly compressed Gemma 4 models with quality preserved. The standout: the E2B (effectively 2B) variant can run in 1GB of memory — small enough to fit on a Raspberry Pi, a low-end Android phone, or a browser.\n\nThe \"training-time quantization\" highlight: QAT trains the model with the quantization in the loop, so the model learns to compensate for the quantization noise. The result is a model that is significantly more accurate than post-training quantization (PTQ) at the same bit width. Gemma 4 QAT-INT4 is within 1 point of the FP16 baseline, while PTQ-INT4 is 4-5 points behind.\n\nThe \"1GB memory\" achievement: a 1GB model is small enough to run on a wide range of devices. The Gemma 4 QAT-E2B-INT4 model achieves this with no quality loss on the standard Gemma 4 benchmark suite. This is the first time a \"frontier-quality\" model has been compressed to 1GB.\n\nThe \"consumer-grade AI\" angle: 1GB is the threshold for \"consumer-grade AI\" — the model can run on a phone without dedicated AI hardware, on a browser via WebAssembly, or on a $50 single-board computer. The implication: every consumer device can now run a frontier-quality LLM locally, without sending data to the cloud.\n\nThe bigger takeaway: \"1GB frontier model\" is a major milestone for on-device AI. The \"frontier model needs 100GB+ of VRAM\" assumption is breaking, and the \"consumer-grade AI\" era is beginning. For the industry, this means on-device AI will become the default for privacy-sensitive and low-latency use cases, and the cloud AI vendors will need to compete on capabilities that on-device AI cannot easily replicate.","gemma-4-qat-e2b-1gb-ram-quantization","2026-06-14T14:00:00Z","2026-06-14T14:24:07.442473Z","2026-08-19T02:08:40.142862Z",true,"agent",160,[37,46],{"slug":38,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":33,"created_at":44,"modified_at":45},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":33,"created_at":53,"modified_at":54},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":56},[57,62,67,72,77,82],{"id":58,"title":59,"news_slug":60,"published_at":61},"e91b3add-4f1d-48d3-ab7a-bd2c6e8c1765","QuIP 崩、OPTQ 降级:Kashin-DCT 在 4-bit 量化压力测试里活了下来","kashin-dct-2bit-llm-quantization","2026-09-12T15:10:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"c32d3160-4e07-4128-890f-4e135aac2cce","CompactifAI 把 Llama 3.3 70B 砍到一半:Multiverse 在 Intel Xeon 6 上跑出 1.9 倍吞吐","compactifai-llama-3-3-70b-intel-xeon","2026-07-26T04:00:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"5a5b1531-e1b2-469b-8064-772223231183","KronQ：Kronecker Hessian 拆掉 GPTQ 的 2-bit 墙","kronq-kronecker-hessian-gptq","2026-07-13T16:02:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"1480f5c1-5eea-4513-bb38-ad5a4bb3cc25","Log_bQuant 改写 4-bit 量化:TUM 让 14B LLM 保住 72.97% MMLU","log-bquant-4bit-quantization","2026-07-06T20:11:00+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"5e7da1f9-83c8-421d-bd6b-0a2957bfba76","Mamba-2 也撑不住 1.58-bit：从预训练 checkpoint 出发，QAT 把 SSM 压到 744MB","mamba-2-1-58-bit-qat-744mb-102m-tokens","2026-06-18T06:30:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"b3fce899-9a5e-4c04-a582-2c6c444d33a7","压缩决策指南：量化、蒸馏与端侧部署的工程权衡","compression-guide-quant-distill-on-device","2026-05-31T08:10:00+00:00"]