Keep on to blur preview images; turn off to show them clearly
什么才叫便宜大碗模型啊，战术后仰。 本地部署大模型的福音来了！给大家带来月之暗面刚发布的 Kimi-Linear-48B-A3B 的技术解析！ 先来一句话版本——这才是便宜大碗快餐模型。 48B-A3B 这个水平做到了1M上下文，然后还是线性注意力，非常省内存。传统注意力上下文长度增长带来的内存消耗是指数级的，这个是线性的，所以这个模型在CPU跑都没事。我已经正在下载了，准备加入本地常用模型中。 目前最大的不确定性是不知道召回水平咋样，我准备下载下来给它塞几本小说问问小说细节看看模型回答的咋样来评估召回效果。想看结果的各位精神股东请点赞，超过100给大家周末放出测评。
基础参数
RT @Sumanth_077: Turn complex documents into RAG-ready data! Agentic Document Extraction (ADE) lets you convert visually complex documents…
专注 - Context Engineering, AI(Coding)Agents. 分享 - AI papers, apps and OSS. ex Microsoft MVP 合作 - 私信/邮箱：shaomeng@outlook.com 📢 公众号/小红书: AI 启蒙小伙伴 🔗 信息卡提示词 🔽
哈哈哈， 画面感太强烈了
Growth Coach｜Helping creators build their personal brand on X 增长教练｜帮助创作者在X上打造个人品牌 公众号：PandaTalk8 X 增长群群主
A protocol can either be: 1⃣ autonomous, credibly neutral and censorship resistant; or 2⃣ capable of dynamic risk scoring to combat illicit finance. It cannot be both because Option 2 requires delegating gatekeeping powers to an oracle. Option 1 is DeFi. Option 2 is not.
@a16zcrypto | Prev Partner @Lathamwatkins | Write about crypto policy, decentralization, tokens & more - https://t.co/HHs2dV36CB
RT @HuggingPapers: Tencent just released a powerful, quantized DeepSeek-V3.1-Terminus model on Hugging Face. DeepSeek-V3.1-Terminus-W4AFP8…
AI research paper tweets, ML @Gradio (acq. by @HuggingFace 🤗) dm for promo ,submit papers here: https://t.co/UzmYN5YmrQ
Always love the millions of ways we find to trick our biological eyes & our visual processing.
Founder @oddtalesgames Directing The Last Night @TLN_Game Art Direction, Cinematography, Tech Art. Atoms, Bits, Memes, Genes. Freedom, Futurism, Humanism.