← 所有主題指南← ALL TOPIC GUIDES
開源模型與開放權重Open Models & Open Weights
全站開源模型文章總覽:開放權重發布解析、本地部署現實、開源與前沿模型的差距,以及圍繞開放權重的政策辯論。Every post here on open models: open-weight release breakdowns, the realities of local deployment, the gap with closed frontier models, and the policy debate around open weights.
主題簡介ABOUT THIS TOPIC
這個主題收集全站與開源模型、開放權重相關的文章:GLM、Qwen、Kimi、DeepSeek、Mistral 等模型發布時的架構與授權解析,以及「權重下載了之後」的本地部署現實。
除了發布消息,這裡也收錄對開放權重生態的結構性分析:開源模型與前沿的差距、各實驗室對開放權重的立場,與圍繞禁令的政策辯論。
必讀精選優先收錄含實測與架構拆解的文章;時間線可以回看每次重要發布。
This hub collects the coverage of open models and open weights on this site: architecture and license breakdowns when GLM, Qwen, Kimi, DeepSeek, Mistral, and others ship, and the practical realities of running those weights locally once downloaded.
Beyond release notes, it includes structural analysis of the open-weight ecosystem: how far open models trail the frontier, where major labs stand, and the policy debate over restrictions.
The must-read picks favor posts with benchmarks and architecture teardowns; the timeline lets you revisit each major release.
必讀精選MUST READ
6GLM-5.3 權重開放下載:753B 旗艦的本地部署現實
2026 年 8 月 25 日,Z.ai 把 753B 參數的 GLM-5.3 權重放上 Hugging Face,三天後在 Hacker News 衝上 806 分。本文解析其 MoE 架構、自訂授權條款與本地部署實測。
閱讀 ↗Kimi K3 登場:2.8T 參數、百萬 Token Context,開源模型走向長程 Agent
Moonshot 發布 Kimi K3:2.8T 參數的開源 3T 級模型,原生視覺、1M token context,主打長程 coding 與知識工作。本文整理 Delta Attention 架構、kernel 編譯器與晶片設計等案例、API 定價與 64 卡部署建議,以及官方自認仍落後 Fable 5 與 GPT 5.6 Sol 的誠實定位。
閱讀 ↗Mistral 的 30 億歐元賭注:開放權重如何成為主權 AI 的技術前沿
Mistral 完成歐洲科技業最大規模融資,以開放權重模型、基礎設施與產品全端布局,回應企業與政府對主權 AI 的需求。本文從產品建造者角度,解析這輪融資對技術選擇與部署策略的意涵。
閱讀 ↗Epoch AI:中國模型平均落後美國前沿 7 個月,差距 4 到 14 個月
Epoch AI 於 2026 年 1 月 2 日發布 ECI 分析:2023 年以來中國模型平均落後美國前沿 7 個月,區間 4 至 14 個月,尚無模型超越 o3。本文拆解計算方法、開源權重的干擾變數與社群爭論。
閱讀 ↗Dario Amodei 表態:Anthropic 從未主張禁用開放權重模型
美國官員傳出考慮禁用中國開放權重模型、科技業連署力挺開放權重之際,Amodei 於 7 月 27 日撰文澄清 Anthropic 從未主張禁令,改提出晶片出口管制、打擊工業級蒸餾與強制安全測試三項主張。
閱讀 ↗Meta 回歸開放權重:30B 的 Muse Glimmer 把本地代理變成現實
2026 年 8 月 10 日,Meta 以 Apache 2.0 釋出 30B 參數的 Muse Glimmer,針對常駐本地代理工作流設計,單張消費級 GPU 就能跑,並承諾開放 Muse Spark 權重。
閱讀 ↗
GLM-5.3 Weights Are Out: Running a 753B MoE Model Locally
Z.ai put the 753B-parameter GLM-5.3 weights on Hugging Face on August 25, and the Hacker News thread hit 806 points. The architecture, the custom license, and local-run realities.
READ ↗Kimi K3 Arrives: 2.8T Parameters, Million-Token Context, Open Models Go Long-Horizon
Kimi K3: a 2.8T open 3T-class model with native vision and 1M-token context for long-horizon coding. The architecture, kernel and chip cases, pricing, and the gap to Fable 5 and GPT 5.6 Sol.
READ ↗What a €3B Sovereign AI Bet Means for Builders
Mistral's Series D signals a shift from raw model power to control over data, models, compute, and production systems. Here's what that means for teams choosing AI infrastructure.
READ ↗Epoch AI: Chinese Models Trail US Frontier by Seven Months
Epoch AI's ECI index puts Chinese models seven months behind the US frontier on average since 2023, ranging four to fourteen months. How it is measured, and the open-weight catch.
READ ↗Amodei: Anthropic Never Sought an Open-Weights Ban
Amodei says Anthropic never advocated a ban on open-weights models, and instead backs chip export controls, a distillation crackdown, and mandatory safety testing.
READ ↗Meta's Muse Glimmer: a 30B Open Model for Local Agents
Meta released Muse Glimmer on August 10, 2026: a 30B open-weight model under Apache 2.0, distilled for always-on local agent workflows on a single consumer GPU.
READ ↗
完整時間線
120 篇文章展開全部 120 篇文章
2026 / 117 篇
- 當瀏覽器內建 AI:Mistral 與 Mozilla 把主權與隱私放進 Firefox Smart Window
- Hermes Agent:記憶留在本機、會自己長出技能的開源助理
- 當實驗數據多到看不完:SAM 3 與 DINOv3 在 Genesis Mission 裡的角色
- 把翻譯模型放進產品前,先看吞吐量與長文件的真實落差
- Mistral 的 30 億歐元賭注:開放權重如何成為主權 AI 的技術前沿
- Ox Alpha 就是 GLM-5.3-Flash:MIT 開源,OpenRouter 週流量佔 31%
- Mistral Forge:不微調、不 RAG,從零訓練企業專屬模型的另類路線
- 加州年齡驗證法修法豁免開源系統:AB 1856 解讀
- Nvidia 洽購 Hugging Face:逾 130 億美元的開源樞紐爭奪
- GLM-5.3 權重開放下載:753B 旗艦的本地部署現實
- Haiku R1/beta6 出爐:25 週年後的回歸之作
- AWS 收購 DuckLabs:DuckDB 母公司加入 Amazon,開源授權不變
- Qwen3.8-Flash-Next 發表:新架構把啟用參數壓到 6B
- DuckDB v2.0 預覽:新解析器、新儲存格式與伺服器模式
- Nari Labs 把 Qwen3-TTS 壓進 50 毫秒內開口
- Go 1.27 登場:泛型方法、json v2 與後量子加密
- Qwen3.8-27B 開源釋出:體積小、看得懂圖,只是預設想太多
- DeepSeek 開源 Agent Harness:一切皆外掛的開發者預覽
- Firefox 成最後防線:uBlock Origin 的 MV2 保衛戰
- Z.ai 發表 GLM-5.3:開源程式碼新高,網路攻擊能力超預期
- Mojo 1.0 正式登場:給 AI 時代的穩定系統語言
- 美國能源部啟動 Genesis 開放模型計畫:開放權重做科學
- Meta 回歸開放權重:30B 的 Muse Glimmer 把本地代理變成現實
- Oracle 劃紅線:AI 生成的內容不得進入 OpenJDK
- Mistral 開源 Shieldstral:把審查政策變成一句提問的 3B 分類器
- MiniMax H3 開源:一次生成 2K 影片與立體聲的全模態模型
- FFmpeg 9.0「Lei」發布:swscale 重寫與更安全的預設
- Qwen3.8-Max 正式發布並開放權重:鎖定寫程式與代理協作
- 在 8GB Mac 上跑 26B 模型:TurboFieldfare 的 SSD 串流推理
- DeepSeek V4-Flash 0731:只重做後訓練,代理能力越級跳
- Dario Amodei 表態:Anthropic 從未主張禁用開放權重模型
- Keychron 開源滑鼠韌體 ZGM:把 QMK 的開源文化帶進遊戲滑鼠
- Debian 公開決議:LLM 貢獻的去留之爭
- 印度下令 GitHub 下架 Bitchat:藍牙通訊遇上審查
- 機場一組密碼清空手機:GrapheneOS 用戶遭聯邦起訴
- Ruff 0.16:預設規則從59條拉到413條
- FLUX 3 登場:影片、圖像、聲音一個骨幹全包
- 輝達、微軟、Meta 公開信:別過早設限開放權重模型
- 近 200 家新創連署:別封殺中國開源權重模型
- Block 開源 Buzz:團隊聊天、AI 代理與 Git 託管共用一條事件流
- Firefox 153 推出:Vulkan 影片解碼與 JPEG XL 終於上船
- Kimi K3 登場:2.8T 參數、百萬 Token Context,開源模型走向長程 Agent
- LM Studio 推出 Bionic:開源模型專用的本機 AI Agent
- Grok Build 開源:Coding Agent 最值得讀的是 Harness,不是 UI
- Reflection 與 Nebius 簽 10 億美元算力合約
- LangChain 攜 NVIDIA 推 NemoClaw:治理優先的 Deep Agents 藍圖
- Bun 1.4 改寫成 Rust:64 個 Claude 代理 11 天完成的大搬遷
- Mistral 開源 Leanstral 1.5:miniF2F 滿分、每題 4 美元的 Lean 證明
- Meta Brain2Qwerty v2:腦波轉文字達 61%
- Ai2 拆解混合模型優勢:贏在內容詞,輸在複製
- Patch the Planet:用 AI 幫開源維護者補洞,而不是增加負擔
- GLM-5.2 開源發布:百萬上下文長任務表現緊咬 Opus 4.8
- Baseten 傳以 130 億美元估值募 15 億美元:五個月估值跳 160%
- Sarvam AI 估值 15 億美元:HCLTech 領投下的印度 AI 獨角獸
- Vercel 開源 eve:agent 框架終於有了一個標準形狀
- Mistral 傳募 30 億歐元、估值 200 億:歐洲主權 AI 的加注與鴻溝
- Google 開源 DiffusionGemma:擴散式生成快 4 倍,單卡 H100 破千 token/秒
- TensorZero 收攤:730 萬美元種子輪的開源 LLM 閘道一夜封存
- Hello Robot Stretch 4 開賣:3 萬美元家用機器人的務實路線
- Miasma 蠕蟲再襲 Microsoft:73 個儲存庫停用,AI 編碼代理成靶
- 微軟開源 ASSERT:把文字規格變成 AI 行為測試套件
- Liquid AI LFM2.5-8B-A1B 發布:38T tokens 訓練的端側 MoE 推理模型
- DeepSeek 把 V4 Pro 七五折降價常態化:前沿模型價格戰的底牌
- Runtime(YC P26)上線:把沙盒化 coding agents 開放給整個團隊
- Stable Audio 3.0 發表:開放權重、6 分 20 秒完整作曲
- NanoClaw 走紅之後:拒絕 2,000 萬美元收購,把 agent 關進容器
- Google ERA 登上 Nature:Gemini 寫出專家級科學程式碼
- NHS 畏懼 AI 漏洞挖掘大舉關閉開源庫,GDS 發布指引唱反調
- DuckDB 推出 Quack 協定:內嵌資料庫補上主從架構缺口
- TanStack npm 供應鏈攻擊解析:三個漏洞串出 84 個惡意版本
- Config 募 2,700 萬美元種子輪:韓國製造業押注機器人資料
- DeepSeek 首度對外募資:估值傳從 200 億美元跳到 450 億
- Mini Shai-Hulud 蠕蟲襲捲 npm:連 Mistral SDK 與 SLSA 證明都淪陷
- Redis 之父 antirez 開源 ds4:跑得動前沿模型的本地推理引擎
- SAP 收購 Prior Labs:四年 11 億歐元打造結構化資料 AI 實驗室
- Ghostty 宣布撤離 GitHub:一本停機日誌終結 18 年依賴
- IBM Granite 4.1 登場:8B 稠密模型追平 32B MoE 的開源算盤
- DeepSeek V4 預覽上線:宣稱追平前沿模型,開源陣營再掀波
- 同一顆模型、兩倍差距:四款 CLI 編碼 Agent 腳手架實測
- OpenAI 開源 Privacy Filter:1.5B 參數的 PII 偵測過濾模型
- Android CLI 與官方 Skills:Google 把代理開發帶進終端機
- Qwen3.6-35B-A3B 開源釋出:3B 啟動參數的代理編碼模型
- 內省式擴散語言模型 I-DLM:首次追平同規模自回歸模型
- NVD 棄守 CVE 積壓:AI 讓漏洞洪流沖垮人工管線
- Ai2 開源 WildDet3D:用單張照片預測 3D 偵測框
- Liquid AI 推出 LFM2.5-VL-450M:跑得進手機的 450M 視覺語言模型
- A2A 協定滿一週年:150 家組織、五種 SDK 與 v1.0 穩定規格
- Gemma 4 開源發布:Apache 2.0、MoE 與 256K 上下文
- PrismML 推出 1-bit Bonsai:把 8B 模型壓進 1.15 GB 的端側 LLM
- Meta 開源 TRIBE v2:預測大腦如何回應影像、聲音與語言
- Cohere 開源 Transcribe 語音模型:5.42% WER 登頂 ASR 排行榜
- Cursor Composer 2 被抓包以 Kimi K2.5 為底:開源權重的署名難題
- USCC 報告:中國開源 AI 的「雙循環」正在強化工業主導地位
- 樂天開源 Rakuten AI 3.0:6,710 億參數、GENIAC 打造的日本最大模型
- Mistral 開源 Leanstral:專為 Lean 4 證明工程打造的 120B 稀疏模型
- LeCun 的 AMI Labs 籌得 10.3 億美元種子輪:押注世界模型
- NVIDIA 開源 Nemotron 3 Super:120B 混合 MoE 模型瞄準 Agent 推論吞吐
- Qwen3.5 小模型補齊戰線:9B 在多項基準超越 gpt-oss-120b
- Perplexity 開源 pplx-embed:擴散預訓練打造網頁級檢索嵌入模型
- Sakana AI 發布 Doc-to-LoRA 與 Text-to-LoRA:一次前向傳遞生成 LoRA 配接器
- Guide Labs 開源 Steerling-8B:把可解釋性做進模型本身
- 阿里巴巴開源 Qwen3.5:397B 參數、201 種語言的 Agent 時代模型
- Cohere Tiny Aya:33.5 億參數、70+ 語言的本機開源模型
- 智利領軍推出 Latam-GPT:拉美首個本土開源大模型
- MiniMax 開源 M2.5 與 Lightning:SWE-Bench 80.2%,十分之一價格逼近 Opus 4.6
- 智譜開源 GLM-5:從 vibe coding 到 agentic engineering
- Microsoft 新掃描法:不必知道觸發詞,也能抓出 LLM 裡的臥底後門
- Mistral 開源 Voxtral Transcribe 2:即時轉錄挑戰雲端大廠
- 從 Clawdbot 到 OpenClaw:爆紅開源代理一週二改名,安全疑慮升高
- AI 一次找齊 OpenSSL 全部 12 個零日漏洞:資安研究的分水嶺
- Kimi K2.5 開源釋出:原生視覺加上 Agent Swarm 多代理協作
- NVIDIA Earth-2 開放氣象模型全家桶:整條 AI 天氣預報管線開源
- Overworld 開源 Waypoint-1:鍵盤滑鼠即時操控的擴散世界模型
- 憶阻器訓練新法 EaPU:AI 訓練能耗比 GPU 低近百萬倍
- Ultralytics 推出 YOLO26:去 NMS、更快的邊緣視覺模型
- 智譜港股掛牌:中國首家純大模型公司上市,首日收漲 13.2%
- Epoch AI:中國模型平均落後美國前沿 7 個月,差距 4 到 14 個月
2025 / 3 篇
FULL TIMELINE
120 ARTICLESOPEN ALL 120 POSTS
2026 / 117 POSTS
- Mistral Powers Firefox's Smart Window: What Zero Data Retention Changes for Browser AI
- Hermes Agent: Self-Hosted AI That Writes Its Own Skills
- SAM 3 and DINOv3 Cut Beamline Segmentation From a Month to 15 Minutes
- North Small Translate: What 16k Context and 1.4x Throughput Change for Translation Pipelines
- What a €3B Sovereign AI Bet Means for Builders
- Ox Alpha Was GLM-5.3-Flash: 31% of OpenRouter Traffic
- Mistral Forge: No Fine-Tuning, No RAG — Training Enterprise Models From Scratch
- California Exempts Linux from Age-Verification Law
- Nvidia in Talks to Buy Hugging Face for Over $13 Billion
- GLM-5.3 Weights Are Out: Running a 753B MoE Model Locally
- Haiku R1/beta6 Arrives: Two Years of Work After Beta5
- AWS acquires DuckLabs: DuckDB stays open source under MIT
- Qwen3.8-Flash-Next: 125B MoE with only 6B active parameters
- DuckDB v2.0 Preview: Server Mode, New Parser, New Storage
- Nari Labs Gets Qwen3-TTS Talking in Under 50 ms
- Go 1.27 Lands: Generic Methods, json/v2, Post-Quantum TLS
- Qwen3.8-27B: Great Open Weights That Overthink by Default
- DeepSeek Open-Sources Agent Harness in Developer Preview
- Firefox, Last Major Browser Still Supporting uBlock Origin
- GLM-5.3: Open-Weight Coding Frontier With Sharp Cyber Gains
- Mojo 1.0: A Stable Foundation for AI Systems Programming
- DOE Launches Genesis Open Models for Open-Weight Science
- Meta's Muse Glimmer: a 30B Open Model for Local Agents
- Oracle Bans AI-Generated Code from OpenJDK Contributions
- Mistral's Shieldstral: A 3B Open-Weights Safety Classifier
- MiniMax H3 Goes Open With 2K Video and Native Stereo Audio
- FFmpeg 9.0 'Lei': swscale Rewrite and Safer Defaults
- Qwen3.8-Max Goes GA: 1M Context and Open Weights
- Running Gemma 4 26B in 2 GB of RAM: TurboFieldfare
- DeepSeek V4-Flash 0731: Same Architecture, Sharper Agents
- Amodei: Anthropic Never Sought an Open-Weights Ban
- Keychron's ZGM Brings Open-Source Firmware to Gaming Mice
- Debian's General Resolution on LLM Contributions
- India Orders GitHub to Take Down Dorsey's Bitchat
- GrapheneOS Duress PIN Wipe Leads to Federal Prosecution
- Ruff 0.16 Ships 413 Default Rules, Up From 59
- FLUX 3: One Backbone for Video, Images, Audio — and Robots
- Nvidia, Microsoft, Meta Warn Against Open-Weight Curbs
- Startups Ask Trump Not to Cut Off Chinese Open-Weight AI
- Block Open-Sources Buzz: Chat, AI Agents, Git in One Place
- Firefox 153 Ships Vulkan Video Decoding and JPEG XL
- Kimi K3 Arrives: 2.8T Parameters, Million-Token Context, Open Models Go Long-Horizon
- LM Studio Bionic: A Local-First AI Agent for Open Models
- Grok Build Open Source: The Harness Is the Part Worth Reading, Not the UI
- Reflection Signs $1B Compute Deal With Nebius
- NemoClaw: LangChain and NVIDIA's Governed Deep Agents Stack
- Bun 1.4 Is Rust Now: 64 Claude Agents Ported It in 11 Days
- Leanstral 1.5: Mistral Saturates miniF2F at $4 a Proof
- Meta Brain2Qwerty v2: 61% Word Accuracy Without Surgery
- Ai2 Maps Where Hybrid LLMs Beat Transformers, Token by Token
- Patch the Planet: How OpenAI and Trail of Bits Are Using AI to Help Open Source Maintainers
- GLM-5.2: Open Weights, 1M Context, Long-Horizon Gains
- Baseten's $1.5B Round Bets Big on Open-Source Inference
- Sarvam AI Turns Unicorn as HCLTech Leads $234M Round
- eve: Vercel's Open-Source Agent Framework Finally Gives Agents a Standard Shape
- Mistral Rumored to Raise €3B at €20B Valuation
- Google Open-Sources DiffusionGemma: 4x Faster Generation
- TensorZero Archives Repo After $7.3M Seed, Winds Down
- Hello Robot Stretch 4: A $30,000 Bet on Home Robots
- Miasma Worm Hits Microsoft Repos, Targeting AI Coding Agents
- Microsoft ASSERT Turns Text Specs into AI Behavior Tests
- Liquid AI LFM2.5-8B-A1B: On-Device MoE Reasoning Model
- DeepSeek Makes Its 75% V4 Pro Price Cut Permanent
- Runtime (YC P26): Sandboxed Coding Agents for Whole Teams
- Stable Audio 3.0: Open-Weight Models, Six-Minute Songs
- NanoClaw Turned Down $20M to Keep Building Sandboxed Agents
- Google's ERA Lands in Nature: Expert-Level Scientific Code
- NHS Retreats from Open Source; GDS Says Keep Code Open
- DuckDB Ships Quack, Its Own Client-Server Protocol
- Inside the TanStack npm Supply-Chain Compromise
- Config Raises $27M to Be the TSMC of Robot Training Data
- DeepSeek's First Outside Round: Valuation Talk Hits $45B
- Shai-Hulud npm Worm Hits Mistral SDK; Provenance No Defense
- antirez's ds4: A Local LLM Inference Engine Built for Metal
- SAP Buys Prior Labs: €1B Bet on Tabular Foundation Models
- Ghostty Quits GitHub After a Month of Daily Outages
- IBM Granite 4.1: An 8B Dense Model Matching a 32B MoE
- DeepSeek V4 Arrives in Preview: Claiming Frontier Parity, Rattling Open Source Again
- Same Model, 2x Gap: Benchmarking Four CLI Coding Agents
- OpenAI Open-Sources Privacy Filter for PII Detection
- Android CLI: Google's Terminal-First Bet on Coding Agents
- Qwen3.6-35B-A3B: Open-Weight MoE Punches at Agentic Coding
- I-DLM: A Diffusion LLM That Matches Same-Scale AR Quality
- NVD Gives Up on CVE Backlog as AI Inflow Accelerates
- Ai2 WildDet3D: Open 3D Detection from a Single Photo
- Liquid AI Ships LFM2.5-VL-450M, a 450M Edge VLM
- A2A Turns One: 150+ Orgs, Five SDKs, Stable v1.0 Spec
- Gemma 4 Ships Under Apache 2.0: Google's Open Model Reset
- PrismML 1-bit Bonsai: An 8B LLM Squeezed Into 1.15 GB
- Meta's TRIBE v2 Predicts Brain Responses Like a Digital Twin
- Cohere Open-Sources Transcribe, Tops ASR Leaderboard
- Cursor Composer 2 Caught Building on Kimi K2.5 Weights
- USCC: China's Open-Source AI Reinforces Industrial Power
- Rakuten AI 3.0: Japan's Largest Model Goes Open-Weight
- Mistral Open-Sources Leanstral, a Lean 4 Proof Agent
- LeCun's AMI Labs Raises $1.03B Seed to Bet on World Models
- NVIDIA Open-Sources Nemotron 3 Super, a 120B MoE for Agents
- Qwen3.5 Small Models: 9B Rivals gpt-oss-120b at the Edge
- Perplexity Open-Sources pplx-embed Retrieval Models
- Sakana AI's Doc-to-LoRA: Documents Become LoRA in One Pass
- Steerling-8B: Guide Labs' Inherently Interpretable Open LLM
- Alibaba Open-Sources Qwen3.5 for the Agentic AI Era
- Cohere's Tiny Aya: A 3.35B Open Model for 70+ Languages
- Latam-GPT: Latin America's First Homegrown Open-Source LLM
- MiniMax M2.5: Open Coding Weights at a Tenth of Opus Cost
- Zhipu Open-Sources GLM-5: From Vibe Coding to Agentic Engineering
- Microsoft's New Scan Catches Sleeper-Agent LLM Backdoors
- Mistral Voxtral Transcribe 2: Open Speech-to-Text, On-Device
- Clawdbot to OpenClaw: Open-Source Agent Hits Security Wall
- AI Found All 12 OpenSSL Zero-Days in One Release
- Kimi K2.5 Goes Open Source: Native Vision and Agent Swarm Coordination
- NVIDIA Earth-2: First Fully Open AI Weather Stack
- Overworld Open-Sources Waypoint-1, a Real-Time World Model
- EaPU Cuts AI Training Energy Nearly a Million-Fold vs GPUs
- Ultralytics YOLO26: NMS-Free Vision AI Built for the Edge
- Zhipu AI Lists in Hong Kong: China's First Pure-Play LLM IPO
- Epoch AI: Chinese Models Trail US Frontier by Seven Months