OPEN SOURCE OBSERVER
从项目热度走向工程判断：追踪开源工具、模型与开发者生态。
01 / CODE
| 排名 | 项目 | 语言 | 今日新增 | 总 Star |
|---|---|---|---|---|
| 01 | NandhaKishorM/layaNon-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request. | Python | — | 20,243 |
| 02 | browser-use/jev-ultrafastFastest and cheapest web agent | Python | — | 19,001 |
| 03 | eternity4719/HowToLiveBetter按性价比排序的循证生活指南：长寿防病、急救、省钱理财、法律红线、失业与工伤、医保社保、恋爱婚育、怀孕育儿、创业与做平台合规、出国与技能。每条写明成本、收益、证据等级和原始出处，只引期刊论文与官方文件。 | HTML | — | 12,996 |
| 04 | lnkiai/m3e-canvasSketch Material 3 Expressive screens in the browser and turn them into vibe-coding prompts. | TypeScript | — | 8,082 |
| 05 | sapientinc/PRAXISTAutonomous research system for measurable, computer-executable research. | Python | — | 6,716 |
| 06 | tamaratran/fast-jev-compactionClaude Code plugin that replaces the compaction summary with Jev decisions: every tool call and result is scored in one fast request, stale ones are dropped or truncated, everything kept stays verbatim. | TypeScript | — | 6,555 |
| 07 | zai-org/ZCodeZ.ai's coding agent harness. Powerful, intelligent, extensible. | TypeScript | — | 6,525 |
| 08 | XiaoDuoYa/codex-with-chatgptChatGPT thinks. Codex works. Use ChatGPT as the planning brain while keeping the Codex harness. | TypeScript | — | 6,517 |
| 09 | rakanki911/DLSS5-SwapperInstall, tune and restore DLSS 5 in your games with one click. Native RenoDX, DLSS5-Feeder for games without DLSS, OptiScaler and multipass routes; DirectX 8/9/11/12, Vulkan, OpenGL, DirectDraw and emulators. In-game F8 overlay, automatic backups, and a community page showing what works on your games and your graphics card. | JavaScript | — | 6,459 |
| 10 | mizorewww/laya-mlxNative MLX runtime for Laya typed decision models — 7–14 ms short decisions on M3 Max. No text generation, PyTorch, or cloud API. | Python | — | 5,921 |
02 / MODELS
| 排名 | 模型 | 任务 | 收藏 | 下载 |
|---|---|---|---|---|
| 01 | Z-Image-Turbo | 文本生成图片 | 2,542 | 820,359 |
| 02 | MiniMax-H3 | — | 2,026 | 154,622 |
| 03 | Qwen3.8-27B | 视觉多模态理解 | 1,514 | 498,817 |
| 04 | DeepSeek-R1 | — | 1,388 | 927,401 |
| 05 | Qwen3.8-Flash-Next | 视觉多模态理解 | 1,270 | 66,849 |
| 06 | Qwen3-Coder-480B-A35B-Instruct | 文本生成 | 1,095 | 138,513 |
| 07 | Qwen-Image-2.1 | 文本生成图片 | 996 | 22,966 |
| 08 | IndexTTS-2 | 语音合成 | 987 | 209,939 |
| 09 | chatglm3-6b | — | 830 | 1,590,966 |
| 10 | Qwen3.8-2.4T-A95B | 文本生成 | 774 | 2,060 |
DEVELOPER VIEW
Laya 系最实用：无文本生成、无 PyTorch、无云 API，可以直接嵌进 agent 循环做路由和缓存命中判断。fast-jev-compaction 的接入成本极低——它就是一个插件，解决的是长会话丢上下文这个高频真实痛点。codex-with-chatgpt 则展示了另一种信号：harness 正在被当成可替换组件，ChatGPT 当规划脑、Codex 保留执行壳，二者可以拆开换。m3e-canvas 把 Material 3 Expressive 草图直接转成 vibe-coding 提示词，适合前端快速起步。HowToLiveBetter 虽非 AI 项目，但"成本/收益/证据等级/原始出处"的结构化写法，本身就是值得借鉴的知识组织范式。
WORTH A STAR
1. NandhaKishorM/laya + mizorewww/laya-mlx：新范式代表。把决策从生成中剥离，本地毫秒级，值得评估是否能用在你现有 agent 的每一步判断上。
2. tamaratran/fast-jev-compaction：最小改动、最大收益的一类工具。如果你的 Claude Code 会话经常被 compaction 截断丢信息，这个插件直接对症。
3. XiaoDuoYa/codex-with-chatgpt：想看清"规划与执行解耦"到底怎么落地，这是当下最直白的样本。
SOURCES & METHOD
GitHub 搜索 API（近 30 天新建 · 按 Star 排序）（项目热度、语言、Star） · ModelScope（模型任务、收藏与下载）
抓取日期：2026-09-24。排名是当日快照，指标可能随平台更新而变化。