科技看板

GitHub · Hugging Face · Ollama · 包下载量 —— 比的是 24 小时增量,不是绝对数

数据 50m前
424h 内发布
07 天新入库
15实验室 7 天新发
0/25有 24h 基线

蜂群解读

主编 + 三席 · 2026-08-29T15:31Z

今日解读

今天有两件事值得注意,都不是星数最高的那个。

Gemini CLI 在配置层对 MCP 做权限隔离。 v0.59.0-nightly 在受限信任模式下将 mcpServerspolicySettings 中完全剥离,这是启动时的 fail-closed 策略,不是运行时的逐条确认。这说明 MCP 在 Gemini CLI 的架构里已经从实验性功能变成了需要安全模型核心介入的组件。我放弃了 OpenAI Codex(0.152.0-alpha.1,4 小时前发布,周下载 1943 万),因为它的 release 页面返回 404,没有可验证的变更日志——有数字没内容,是噪音。

GLM-5.3 家族正在快速穿透推理栈。 Z.ai 四天前在 Hugging Face 发布 GLM-5.3 和 GLM-5.3-Flash,Ollama 在 24 小时内完成收录,Unsloth 的 GGUF 量化版也已上线。这是 base model → 分发 → 量化的完整链条。但要注意:Ollama 上只有 Flash 版本带 vision 标签,旗舰版 GLM-5.3 的标签是 tools thinking cloud,没有 vision。我也放弃了 tt-a1i/archify(今日 +3927 星),因为 agent skill demo 的 trending spike 通常是社交放大,而我们没有 24 小时基线来判断是否持续。

我可能错在哪:Gemini CLI 的变更只限于 a2a-server 子包,它是否默认启用、影响面多大,目前不清楚。GLM-5.3 的 adoption chain 看起来干净,但没有 24 小时基线,无法确认 Flash 的 22,800 次拉取是真实速度还是首发前置。

席位发现

- agent_stack_watch:MCP 隔离发生在配置加载阶段,而非调用阶段。当 trusted = false 时,safeMcpServers 被设为 undefined 后再写入 policySettingsconfigParams。这对在受限模式下依赖仓库级 .gemini/settings.jsonmcpServers 的用户是 breaking change,但 PR 未明确标记为 breaking。变更仅影响 a2a-server 子包,不触及核心 CLI 的 MCP 调用路径。Codex 对比无法完成——release 页面 404,changelog 重定向到被截断的 releases 页。

- model_release_watch:GLM-5.3-Flash 经 Ollama 的 vision 标签和自述确认为多模态模型,是 "Z.ai 首个原生多模态模型"。GLM-5.3 旗舰版无 vision 标签,定位为纯文本的 coding/agentic 模型。Ollama 拉取速度:Flash 约 7,600/天,qwen3.8-flash-next 约 4,700/天,但两者格式体积不同(qwen 的 125B MLX 为 105GB,可能抑制下载)。NVIDIA 的 DeepSeek-V4-Pro-0813-NVFP4(0 赞 0 下载)使用的 "V4" 命名与 DeepSeek 官方公开版本线(止于 V3/R1)不符,可能是内部实验命名或误标。

变化

ollama/ollama 发布 v0.33.218 小时前发布ggml-org/llama.cpp 发布 b106820 小时前发布OpenHands/OpenHands 发布 v1.16.045 小时前发布openai/codex 发布 0.152.0-alpha.14 小时前发布google-gemini/gemini-cli 发布 Release v0.59.0-night…13 小时前发布langchain-ai/langgraph 发布 langgraph-sdk==0.4.442 小时前发布

GitHub Trending今日新增星

  • tt-a1i/archifyAgent skill for beautiful, verifiable architecture, workflow, sequence, data-flow, and lifecycle diagrams—self-contained HTML with motion and crisp export.
    JavaScript29.9k+3,927
  • bilawalsidhu/gods-eye-viewA spy satellite simulator in your browser, except the data is real. Live open source spatial intelligence on a photorealistic 3D globe.
    JavaScript12.1k+1,870
  • K-Dense-AI/scientific-agent-skillsTurn any AI agent into an AI Scientist. The #1 Agent Skills library for science, used by 190,000+ scientists worldwide. 165 ready-to-use validated skills plus 100+ scientific databases covering biology, chemistry, medicine, and drug discovery. Compatible with Cursor, Claude Code, Codex, Pi, Antigravity, and the open Agent Skills standard.
    Python37.5k+1,604
  • calesthio/OpenMontageWorld's first open-source, agentic video production system. 12 production pipelines, 100+ tools, 700+ agent skill and production-knowledge files. Turn your AI coding assistant into a full video production studio.
    Python53.8k+809
  • tailscale/tailcatlike netcat, but over Tailscale's data plane, without Tailscale's control plane
    Go3,224+790
  • freestylefly/awesome-gpt-image-2Prompt as Code | GPT-Image2 工业级提示词引擎与模板库,530+ 个案例逆向工程,20+ 套工业级模板,并提炼出Skills,持续更新中
    JavaScript24.9k+767
  • abi/screenshot-to-codeDrop in a screenshot and convert it to clean code (HTML/Tailwind/React/Vue)
    Python75.9k+558
  • NationalSecurityAgency/ghidraGhidra is a software reverse engineering (SRE) framework
    Java73.5k+375
  • anthropics/claude-plugins-officialOfficial, Anthropic-managed directory of high quality Claude Code Plugins.
    Python35.3k+356
  • JetBrains/go-modern-guidelinesHelp AI coding agents write modern Go
    Go2,777+294
  • workweave/routerModel router for agentic systems. Routes every prompt to the right model in <50ms. Cut costs 40-70% with just an endpoint change.
    Go2,550+284
  • abhigyanpatwari/GitNexusGitNexus: The Zero-Server Code Intelligence Engine - GitNexus is a client-side knowledge graph creator that runs entirely in your browser. Drop in a git repository (Github, Gitlab, Azure, Local) or ZIP file, and get an interactive knowledge graph with a built in Graph RAG Agent. Perfect for code exploration
    TypeScript46.3k+273
  • cursor/pluginsCursor plugin specification and official plugins
    TypeScript6,097+257
  • vxcontrol/pentagiFully autonomous AI Agents system capable of performing complex penetration testing tasks
    Go22.2k+45
  • vllm-project/semantic-routerA programmable Mixture-of-Models router for heterogeneous LLM inference
    Go5,403+27
  • grafana/alloyOpenTelemetry Collector distribution with programmable pipelines
    Go3,486+6

关注仓stars · Δ24h · 最新发布

Ollama 模型库pulls · Δ24h

  • 最新入库
    • qwen3.8-flash-nextThis experimental preview of the architecture that will underpin Qwen4.
      8,435·
    • glm-5.3-flashZ.ai's first natively multimodal model, approaching Claude Opus 4.8 on coding and agentic benchmarks with just 18B active parameters.
      22.8k·
    • ornith-1.5Chirp Chirp! 🐦 We are introducing Ornith-1.5, a major step toward building foundation models through end-to-end self-improvement.
      9b35b+1124.1k·
    • glm-5.3Z.ai's flagship model and the most capable open-weights model for coding, with major gains on long-horizon agentic tasks.
      7,908·
    • granite4.2IBM Granite Models are a family of enterprise-ready, open foundation models that support multilingual capabilities, coding, retrieval-augmented generation (RAG), tool use, thinking and structured JSON output. Released under Apache 2.0 license.
      3b8b+115.4k·
    • qwen3.8Qwen3.8 delivers substantial gains across coding, professional work, research, and long-horizon agentic tasks.
      27b1.1M·
    • nemotron-3.5-lightningNVIDIA Nemotron 3.5 Lightning is an open 30B mixture-of-experts (MoE) model with 3B active parameters built for always-on agents.
      30b139.8k·
    • muse-glimmerMeta's latest open model built for always-on local agents. 30B parameters, licensed under Apache 2.0 and runs on a single GPU — tuned for tool use, long tasks, and failure recovery.
      30b174.8k·
    • kimi-k3Kimi K3 is an open-weight, native multimodal agentic model and our most capable model to date.
      63.6k·
    • laguna-s-2.1Our most capable model to date, designed for long-horizon work. 70.2% on Terminal-Bench 2.1 at 118B-A8B.
      116.1k·
    • laguna-xs-2.1Laguna XS 2.1 is a 33B total parameter Mixture-of-Experts model with 3B activated parameters per token designed for agentic coding and long-horizon work on a local machine.
      101.2k·
    • ornithA self-improving family of open-source models for agentic coding
      9b35b430.8k·
    • north-mini-code-1.0North Mini Code is Cohere's first model for developers — a 30B Mixture-of-Experts model with 3B active parameters, built for agentic software engineering.
      46.9k·
    • glm-5.2GLM-5.2 is Z.ai’s flagship model for the era of long-horizon tasks.
      335.3k·
    • kimi-k2.7-codeKimi K2.7 Code is Moonshot AI's coding-focused agentic model built upon Kimi K2.6, with substantial improvements on real-world long-horizon coding tasks and roughly 30% lower thinking-token usage.
      227.3k·
    • nemotron-3-ultraNVIDIA Nemotron 3 Ultra is built for high-throughput reasoning and long-running agent workflows.
      64.9k·
  • 热门
    • llama3.1Llama 3.1 is a new state-of-the-art model from Meta available in 8B, 70B and 405B parameter sizes.
      8b70b+1118.9M·
    • deepseek-r1DeepSeek-R1 is a family of open reasoning models with performance approaching that of leading models, such as O3 and Gemini 2.5 Pro.
      1.5b7b+592.1M·
    • nomic-embed-textA high-performing open embedding model with a large token context window.
      83.9M·
    • llama3.2Meta's Llama 3.2 goes small with 1B and 3B models.
      1b3b81.7M·
    • gemma3The current, most capable model that runs on a single GPU.
      270m1b+339.9M·
    • qwen2.5Qwen2.5 models are pretrained on Alibaba's latest large-scale dataset, encompassing up to 18 trillion tokens. The model supports up to 128K tokens and has multilingual support.
      0.5b1.5b+538.8M·
    • qwen3Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models.
      0.6b1.7b+635.7M·
    • mistralThe 7B model released by Mistral AI, updated to version 0.3.
      7b33.1M·
    • gemma2Google Gemma 2 is a high-performing and efficient model available in three sizes: 2B, 9B, and 27B.
      2b9b+131.3M·
    • llama3Meta Llama 3: The most capable openly available LLM to date
      8b70b25.1M·
    • gemma4Gemma 4 models are designed to deliver frontier-level performance at each size. They are well-suited for reasoning, agentic workflows, coding, and multimodal understanding.
      e2be4b+323.7M·
    • qwen2.5-coderThe latest series of Code-Specific Qwen models, with significant improvements in code generation, code reasoning, and code fixing.
      0.5b1.5b+420.8M·
    • qwen3.5Qwen 3.5 is a family of open-source multimodal models that delivers exceptional utility and performance.
      0.8b2b+518.7M·
    • phi3Phi-3 is a family of lightweight 3B (Mini) and 14B (Medium) state-of-the-art open models by Microsoft.
      3.8b14b18.1M·
    • llava🌋 LLaVA is a novel end-to-end trained large multimodal model that combines a vision encoder and Vicuna for general-purpose visual and language understanding. Updated to version 1.6.
      7b13b+114.7M·
    • mxbai-embed-largeState-of-the-art large embedding model from mixedbread.ai
      335m14.1M·

Hugging Facetrending · ❤ · Δ24h

包下载量每周 · Δ24h