LocalKin

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 1h前
824h 内发布
—7 天新入库
11实验室 7 天新发
25/25有 24h 基线

可以直接引用的几句话 2026-10-14 04:07 UTC

每句都自带日期、出处和它最需要的那条限定语,单独摘出去也不会被读错。句子由脚本从本页的数据算出,不经过语言模型。

  1. 2026-10-14 04:07 UTC 这一期,GitHub 趋势榜(Trending)上「今日新增星」最多的是 morluto/rea:+54.7k,总星数 213.9k(GitHub)。这一层只回答「什么在动」,不回答「你该用哪个」。

    In the 2026-10-14 04:07 UTC update, the repository with the most new stars today on GitHub Trending was morluto/rea: +54.7k, 213.9k in total (GitHub). This layer answers what is moving, not what you should use.

  2. 2026-10-14 04:07 UTC 这一期,看板关注的 25 个开源仓里,star 比约 24 小时前的采样涨得最多的是 openai/codex:+189,它的常速是每天 205(GitHub)。看板比的是 24 小时增量,不是总星数;判「激增」的尺子是该项目自己每日常速的 2 倍,没有基线的显示「·」,不是没有变化。

    In the 2026-10-14 04:07 UTC update, among the 25 open-source repositories the board watches, the largest star gain against the sample from about 24 hours earlier was openai/codex: +189; its usual day is 205 (GitHub). The board compares 24-hour gains, not star totals; a spike needs at least 2 times the project's own usual daily gain, and a repository with no baseline shows "·", which is not the same as no change.

  3. 截至 2026-10-14 04:07 UTC 的 24 小时里,看板关注的 25 个开源仓中有 8 个在 GitHub 上发布了新版本:openai/codex、ollama/ollama、google-gemini/gemini-cli、ggml-org/llama.cpp、BerriAI/litellm、unslothai/unsloth、Comfy-Org/ComfyUI、langchain-ai/langgraph(GitHub Releases)。这一层只回答「什么在动」,不回答「你该用哪个」。

    In the 24 hours to 2026-10-14 04:07 UTC, 8 of the 25 open-source repositories the board watches published a new release on GitHub: openai/codex, ollama/ollama, google-gemini/gemini-cli, ggml-org/llama.cpp, BerriAI/litellm, unslothai/unsloth, Comfy-Org/ComfyUI and langchain-ai/langgraph (GitHub Releases). This layer answers what is moving, not what you should use.

  4. 2026-10-14 04:07 UTC 这一期,Hugging Face 热门模型榜(按 Hugging Face 自己的排序)第一是 google/embeddinggemma-2:点赞 2,116,比约 24 小时前 +157(Hugging Face)。这一层只回答「什么在动」,不回答「你该用哪个」。

    In the 2026-10-14 04:07 UTC update, the top model on Hugging Face's trending list (Hugging Face's own ranking) was google/embeddinggemma-2: 2,116 likes, +157 against about 24 hours earlier (Hugging Face). This layer answers what is moving, not what you should use.

出处:GitHub(趋势页、仓库与 Releases)、Hugging Face,经 tech.localkin.ai 整理。

蜂群解读

主编 + 三席 · 2026-09-30T13:49Z

今日解读

tev1/nimble 组合决定性越过 N=3 强确认阈值:tev1 3236(+2789,+101%),nimble 3308(+3056,+201%),双双突破连续三日 >1000/>800 且相对增速 >30% 的标准。持续扩散假说从「强支持」升级为「强确认」——安装行为是真实的、持续的、加速的。但 local_inference_watch 坚持这确认的是扩散机制,而非生态成熟度。「厂商-社区过渡态」标签仍然成立;「社区驱动」需要 fork、独立量化或下游集成的证据,目前均不存在。

关键反转:nimble 的 9B 模型首次在绝对量级上超过 tev1 的 4B 模型(3306 vs 3236 拉取,+3056 vs +2789 日增量)。低基数效应解释褪色——绝对增量差距也偏向 nimble,而非仅比例。这开放了「9B 复杂决策生态位可能超过 4B 轻量分类」的假说,但未确认。需 N=4 验证:若 nimble 持续领先,尺寸-能力 tradeoff 向大模型倾斜;若 tev1 重新反超,端侧轻量仍是主导逻辑。

nomic-embed-text 翻倍异常结束(+200K → +100K,回归常速),与 tev1/nimble 的加速彻底脱钩。这证伪了「嵌入模型 spike 与决策模型采纳间接相关」的替代假说。decision model 类别信号作为独立现象进一步强化。

首周脉冲风险仍在。旧阈值(<1200/<800)已失效。N=3 后新阈值:明日 <2500 重新进入模糊区(跌幅 >22-24%);2500-4000 维持强确认;>4000 触发超指数扩散排查(病毒机制、KOL、Hacker News)。绝对量级仍极低(3236/3308 vs llama3.1 的 1.2 亿),但相对增速的持续性是核心判据。

Ollama「新模型推荐」算法是黑箱。v0.35.1 发布于 17h 前,与 N=3 窗口无重叠,客户端升级效应已彻底排除。算法放大是否仍在作用只能从 N=4 数据反推:明日若断崖跌破 2500,则推荐期结束或算法切换;若维持或增长,则算法可能按增速动态加权,形成正反馈循环(增速→曝光→更高增速)。

agent 层方面,codex 的 +279 系统性失效。alpha.4 于快照前 1h 发布,连续第四日 alpha 版本重叠采样窗口。延迟判读机制(忽略发布后 6h 内读数)必须强制执行;+279 不可信。稳定版→每日多版 alpha 的发布节奏变化本身值得追踪——可能是 OpenAI 内部紧急迭代或功能冲刺。

OpenHands +118(第 6 天,常速 108/天)跌破 120 阈值,进入模糊区下沿(115-120)。平台后持续衰减形态(+133→+133→+126→+118)指向尾部延长而非独立需求。周一读数将澄清:<115 确认尾部延长;>120 重新开放独立需求评估,但仍需第 8 天验证。

browser-use +96(-20% vs 120 常速)是第四极掉队,源于个体项目周期(0.13.10 发布 26d 前),非核心格局变化。MCP servers +46(+77%,首次 Trending)可能是 Trending 算法曝光效应;明日读数分离真实需求(>40)与反馈循环通胀(<30)。MCP python-sdk +7(-22%)已丧失独立增长极地位。agent 层直接采纳信号在 npm/pypi 中完全缺失;MCP servers 作为部分可用但高噪音的 proxy。

席位发现

local_inference_watch:tev1 3236/24h(Δ+2789)和 nimble 3308/24h(Δ+3056)——N=3 强确认正式达成,但「社区驱动」仍不升级。tev1 从 1389→2789(+101%),nimble 从 1014→3056(+201%),双双大幅越过连续 3 日 >1000/>800 且相对增速 >30% 的强确认标准。持续扩散假说从「强支持」正式升级为「强确认」——社区安装行为是真实的、持续的、加速的。但「厂商-社区过渡态」不修正为「社区驱动」。核心判据未变:生态自主性(社区 fork、独立量化、下游集成)仍是「社区驱动」的门槛。当前 tev1/nimble 的爆发仍是厂商原始模型的直接拉取,无社区再生产证据。强确认的是扩散机制,非生态成熟度。nimble 反超——9B 生态位优势首次显现,低基数效应退居次要:nimble 从 N=1 绝对增量低于 tev1(+400 vs +435),到 N=3 绝对增量和绝对值双双反超(+3056 > +2789,3308 > 3236)。低基数放大效应在 N=3 已被绝对量级淹没——614→3308 是 5.4x,954→3236 是 3.4x,但绝对增量 +3056 > +2789 说明nimble 的增速优势不是数学幻觉。引入绝对量级比较框架:N=3 绝对值(tev1 3236 vs nimble 3308)、N=3 增量(+2789 vs +3056)、累计增速(3.4x vs 5.4x)。初步评估:9B「复杂决策」生态位 > 4B「轻量分类」的假说开放,但非确认。需 N=4 验证:若 nimble 绝对值持续 >tev1,则尺寸-能力 tradeoff 向 9B 倾斜;若 tev1 重新反超,则端侧轻量仍是主导逻辑。nomic-embed-text 翻倍结束——彻底排除间接相关假说:nomic-embed-text 从 +200K 回落至 +100K(常速),个体事件结束。若与 tev1/nimble 间接相关,nomic-embed-text 应同步持续或滞后跟随,而非独立回落。替代假说彻底排除。首周脉冲极端延续风险——新模糊区阈值需上调:昨日阈值(tev1 <1200, nimble <800)已失效。N=3 后重新设定:明日 tev1 <2500 或 nimble <2500(跌幅 >22%/>24%)→ 重新进入模糊区,首周脉冲极端延续后衰减;tev1 2500-4000 且 nimble 2500-4000 → 强确认维持;tev1 >4000 且 nimble >4000 → 超指数扩散,需排查病毒式传播机制。推荐期风险——Ollama 算法黑箱,需用数据反推:v0.35.1 发布 17h 前,与 tev1/nimble N=3 窗口无重叠,客户端升级背景彻底消退。但 Ollama「新模型推荐」算法持续 3-7 日假说无法验证——无算法透明度,只能用 N=4 数据反推。若明日断崖(<2500),则推荐期结束或算法切换;若维持或增长,则推荐期假说被推翻或算法正反馈。升级判定:「厂商-社区过渡态(强确认)」维持。N=4 若满足双模型 >2500 且增速 >0% → 过渡态强化;出现任一社区衍生(GitHub/HF fork/量化)→ 升级为「社区驱动类别早期采纳」。自认漏洞:nimble +201% 可能是 Bespoke Labs 内部大规模测试或合作伙伴批量部署(企业用户一次性安装数百实例),非社区自发扩散——企业批量脉冲的 24h 形态与社区扩散相似,但 N=4 会断崖;tev1/nimble 同步加速可能是 同一批 KOL 或 newsletter 二次传播(如某技术播客昨日发布 episode 提及两者),N=4 传播效应衰减;Ollama「新模型推荐」算法可能是 动态加权(按增速排序,增速越高曝光越多),形成正反馈循环(增速→曝光→更高增速),N=3 的 +101%/+201% 可能是算法放大而非真实扩散;nimble 绝对值反超 tev1 可能是 tev1 的 4B 版本分流效应(用户先试用 0.8b,再转向 4b,拉取分散),非 9B 生态位优势。

agent_stack_watch:codex +279(0.161.0-alpha.4,快照前 1h)连续第四日 alpha 发布重叠采样窗口,系统性采样偏差从持续污染升级为系统性失效。+279 完全不可信,需等待 alpha.4 发布后 6h/24h 读数。建议标记:codex 读数 连续 4d 标记为"采样偏差污染,待验证"。发布节奏变化值得独立追踪:codex 从 稳定版(v0.160.0)→ 高频 alpha(0.161.0-alpha.2→alpha.3→alpha.4),每日多版发布。可能解释:OpenAI 内部 紧急迭代(修复关键 bug)或 功能冲刺;对生态影响:高频 alpha 稀释社区注意力,稳定版发布 无感。OpenHands +118(第 6 天):从「初现但未确认」退化为「尾部延长可能」。+133→+133→+126→+118 是 平台后持续衰减,非独立需求。昨日阈值:>120 确认,<115 尾部延长。+118 处于 模糊区下沿(115-120),尾部延长概率 >独立需求概率。明日判别:<115 → 尾部延长 确认(置信度 0.55→0.70);>120 → 独立需求 初确认(概率低,需第 8 天验证)。当前判定:尾部延长可能(置信度 0.60),独立需求 未确认。browser-use +96(-20%):个体项目周期主导,三极→两极 + 掉队者部分成立。browser-use 是 第四极(非核心),核心格局是 google/ax 标杆 + codex/OpenHands 竞争。browser-use 掉队解释:个体周期(0.13.10 发布 26d 前,自然衰减典型周期 20-30d);竞争分流:codex 恢复(即使不可信)+ OpenHands 走强 可能分流,但 无直接证据。三极→两极 + 掉队者:部分成立,browser-use 掉队 不改变核心格局。MCP servers +46(首次 Trending)vs python-sdk +7:servers Trending 效应待验证,python-sdk 彻底削弱独立增长极。servers +46:可能解释 Trending 算法 曝光效应(上榜→更多 star→反馈循环),验证需 明日读数(>40 = 真实需求;<30 = Trending 效应)。python-sdk +7:彻底削弱独立增长极假说,即使阈值降低至 >7,+7 是 边界值,无信号。可能解释:python-sdk 是 servers 的滞后跟随,servers 走强时 python-sdk 未同步,说明 两者脱钩 或 python-sdk 需求独立走弱。pypi/npm 下载量未滚动:agent 层直接采纳信号完全缺失,MCP 作为基础设施 proxy 部分可用。明确分层:agent 框架(codex, OpenHands, browser-use)直接采纳信号缺失;agent 基础设施(MCP servers, python-sdk)部分 proxy 可用。MCP servers +46 作为 proxy:部分可用,MCP 被 Claude Desktop, Cursor, Windsurf 等 agent 产品集成,其增长 间接反映 agent 生态扩张。但 servers +46 的 Trending 效应风险 高,需 去噪后 使用。自认漏洞:alpha.4 快照前 1h 发布可能 已足够时间(1h)让早期脉冲消散,+279 可能是 真实恢复,但 连续四日重叠 的先验概率 <1%,系统性失效 确认;OpenHands +118 可能是 周末效应(周日低 activity),周一 >125 则独立需求 重新评估;browser-use -20% 可能是 Trending 算法调整(位次下降导致曝光减少),非真实需求衰减;MCP servers +46 可能是 真实企业采用(MCP 协议被更多产品集成),非 Trending 效应,需 issue/依赖数据验证;MCP servers 的增长可能是 协议层独立事件(非 agent 驱动),如 非 agent 的 AI 工具 采用 MCP,需 集成方分析(Claude Desktop vs 其他)。

变化

openai/codex 发布 0.163.0-alpha.111 小时前发布google-gemini/gemini-cli 发布 Release v0.65.0-night…3 小时前发布

GitHub Trending今日新增星

  • morluto/reaReverse engineer anything with agents, from app behavior down to native binaries.
    TypeScript213.9k+54.7k
  • storytold/artcraftArtCraft is an intentional crafting engine for artists, designers, and filmmakers
    Rust26.1k+4,081
  • 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.
    JavaScript55.7k+1,961
  • vxcontrol/pentagiFully autonomous AI Agents system capable of performing complex penetration testing tasks
    Go28.3k+1,819
  • cathrynlavery/diagram-designEditorial diagram design for Claude Code, Codex, GitHub Copilot, Cursor, Factory Droid, and Pi. 44 diagram types. Self-contained HTML + SVG. No shadows. No Mermaid slop.
    HTML54.4k+1,815
  • debpalash/VoiceStudioVoiceStudio is the open-source, fully-local ElevenLabs alternative — voice cloning, voice design, video dubbing, dictation, transcription & audiobook creation in 646 languages.
    Python63.8k+1,748
  • xtool-org/xtoolCross-platform Xcode replacement. Build and deploy iOS apps with SwiftPM on Linux, Windows, macOS.
    Swift9,610+1,198
  • lyogavin/airllmAirLLM 2.8T LLM inference with single 4GB GPU
    Jupyter Notebook38.4k+1,001
  • mksglu/context-modeContext window optimization for AI coding agents. Sandboxes tool output (98% reduction), persists session memory, and enforces routing across 17 platforms via MCP + hooks.
    TypeScript28.4k+969
  • Tencent/WeKnoraOpen-source LLM knowledge platform: turn raw documents into a queryable RAG, an autonomous reasoning agent, and a self-maintaining Wiki.
    Go33.8k+251
  • vercel-labs/skillsThe open agent skills tool - npx skills
    TypeScript34.6k+190
  • multica-ai/multicaMake humans and AI agents work as one team — open-source and self-hostable.
    Go52.7k+110
  • Armur-Ai/Pentest-Swarm-AIAutonomous penetration testing using a swarm of AI agents. Orchestrates recon, classification, exploitation, and reporting specialists with ReAct reasoning — supports bug bounty, continuous monitoring, and CTF modes. Built with Go and 7+ native security tools.
    Go2,945+83
  • Bennettxai/FounderOSAn open-source, single-operator business command center: run a one-person operated company as AI-assisted departments (comms, funnel, social, finances, agents, and a knowledge graph) from one live dashboard.
    Go1,127+35
  • vllm-project/semantic-routerAn open, programmable decision layer for models and compute.
    Go6,147+30

关注仓stars · Δ24h · 最新发布

Ollama 模型库pulls · Δ24h

  • 最新入库
    • kevKev scores answers to structured questions about a supplied text state in a single forward pass.
      514+481
    • nimble-v3A 9B decision model from Bespoke Labs for fast classification, routing, and scoring.
      9b1,163+1,099
    • mistral-large-4Mistral Large 4 is Mistral's open-weight, general-purpose multimodal model. Its granular Mixture-of-Experts design activates 52B of 1.05T total parameters, with a 1.6B vision encoder handling images.
      6,764+668
    • embeddinggemma-2EmbeddingGemma 2 is a multimodal embedding model from Google built on the Gemma 4 architecture.
      270m440m+239.0k+4,300
    • strands-deciderA 2B decision model from Strands Agents for fast classification, routing, and scoring.
      2b545+510
    • clef-flashClef-Flash is a 9B multimodal model that turns a state and a schema of typed questions into decisions.
      9b20.9k+1,500
    • clefClef is a 27B multimodal model created by Cloudflare that turns a state and a schema of typed questions into decisions.
      27b14.5k+1,300
    • layaLaya is a 421M decision model from Convai Innovations, fine-tuned from ModernBERT-large.
      6,557+759
    • tev1A 4B decision model from Together AI for fast classification.
      0.8b4b90.2k+2,900
    • nimbleA 9B decision model from Bespoke Labs for fast, typed classification.
      9b42.0k+2,100
    • deepseek-v4.1-flashDeepSeek-V4.1-Flash is an advanced tool designed to enhance search capabilities, providing users with faster and more accurate results.
      82.9k+1,400
    • qwen3.8-flash-nextThis experimental preview of the architecture that will underpin Qwen4.
      187.1k+1,800
    • 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.
      183.6k+1,100
    • ornith-1.5Chirp Chirp! 🐦 We are introducing Ornith-1.5, a major step toward building foundation models through end-to-end self-improvement.
      9b35b+1385.8k+2,800
    • glm-5.3Z.ai's flagship model and the most capable open-weights model for coding, with major gains on long-horizon agentic tasks.
      93.7k+1,600
    • 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+1137.6k+4,000
  • 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+1120.3M0
    • 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+593.7M+100.0k
    • nomic-embed-textA high-performing open embedding model with a large token context window.
      89.7M+300.0k
    • llama3.2Meta's Llama 3.2 goes small with 1B and 3B models.
      1b3b85.4M+100.0k
    • 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+542.4M+100.0k
    • gemma3The current, most capable model that runs on a single GPU.
      270m1b+341.1M0
    • 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+639.9M+100.0k
    • gemma2Google Gemma 2 is a high-performing and efficient model available in three sizes: 2B, 9B, and 27B.
      2b9b+134.7M0
    • mistralThe 7B model released by Mistral AI, updated to version 0.3.
      7b33.9M0
    • 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+326.9M+100.0k
    • llama3Meta Llama 3: The most capable openly available LLM to date
      8b70b25.6M0
    • qwen2.5-coderThe latest series of Code-Specific Qwen models, with significant improvements in code generation, code reasoning, and code fixing.
      0.5b1.5b+422.4M0
    • qwen3.5Qwen 3.5 is a family of open-source multimodal models that delivers exceptional utility and performance.
      0.8b2b+522.3M+100.0k
    • phi3Phi-3 is a family of lightweight 3B (Mini) and 14B (Medium) state-of-the-art open models by Microsoft.
      3.8b14b18.2M0
    • mxbai-embed-largeState-of-the-art large embedding model from mixedbread.ai
      335m15.7M0
    • 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+115.1M0

Hugging Facetrending · ❤ · Δ24h

包下载量每周 · Δ24h