Model Radar / 世界大模型雷达

Model Radar is a static, source-backed switchboard for current frontier models, benchmark rankings, official launches, and inferred provider release cadence.

世界大模型雷达是一块可追溯来源的静态观察台，用来切换查看前沿模型排名、官方发布、benchmark 覆盖和企业发行节奏。

Public site / 在线页面:

https://wychuang.github.io/model-radar/

V2 research desk / V2 研究桌面:

https://wychuang.github.io/model-radar/v2.html

Radar observatory / 雷达观测站:

https://wychuang.github.io/model-radar/radar.html

模型斩杀线 / Intelligence × price:

打开 radar.html#value-frontier。把所有具有可比数据的模型放在智能与价格散点图中， 点选模型即可看到更优替代、差价和智能分差。预算与最低智能滑块筛选前沿模型。 判断采用 AA 智力原始值与每百万输出 tokens 的美元单价；更便宜且不弱、或同价更强， 即可淘汰缺乏优势的选择。缺少价格或智能的模型单独列出。

本地研究、具体案例及维护建议见 REVIEW.md。

V2 keeps the same curated snapshot and ranking logic while presenting the current leader, NOW/NEXT signals, evidence labels, and release timing in a responsive evidence workspace. Its main switcher uses the same seven comparable metrics as the radar observatory, while model-specific specialist scores appear only when an exact measured row exists. Both views link directly to each other.

V2 继续使用同一份人工确认快照与排序逻辑，并把当前领先、NOW/NEXT 信号、 证据标签和发布节奏组织为响应式研究桌面。主切换器与雷达观测站共用七项可比指标， 模型专项分数只在存在完全匹配的实测记录时显示。两个视图可直接来回切换。

Radar observatory details / 雷达观测站说明:

http://127.0.0.1:4174/radar.html

The local radar edition restores the original spatial radar-map / model-dots identity. Its horizontal axis follows the selected source-backed benchmark, its vertical axis shows release freshness, and missing measurements move into a visible NO SIGNAL lane with a verification reason. A second five-axis decision profile compares intelligence, preference, output speed, price efficiency, and context coverage. Four use-case modes automatically select the strongest current choice and the most relevant alternative; engineering and agent benchmarks stay visible as separate raw evidence. The visual and interaction rules live in DESIGN.md.

本地雷达版恢复原始 radar-map / model-dots 空间交互。横轴跟随当前选择的可追溯 benchmark，纵轴显示发布新鲜度，缺失测量进入清晰的 NO SIGNAL 轨道。视觉与 详情会说明当前缺失原因。第二层五维选型对比智能、偏好、生成速度、价格效率和上下文， 并按综合、编程 Agent、高吞吐、性价比四种用途自动给出首选与最相关替代。工程与 Agent 专项 benchmark 保留原始分值。视觉与交互规则记录在 DESIGN.md。

What It Shows / 展示内容

• 20 model rows across 12 provider lanes, curated on 2026-09-10; 17 have comparable intelligence and output-price evidence.

• Seven switchable views: Artificial Analysis Intelligence Index, Arena Text, output speed, SWE-Bench Pro, Terminal-Bench, output price, and advertised context.

• Raw values, measurement dates, source links, preliminary/vendor labels, and visible N/A cells where comparable evidence is missing.

• A five-axis decision profile with four automatic use-case recommendations, coverage gates, honest gaps, and provider-family colors.

• Model-specific specialist evidence including AA Coding Agent Index, GDPval-AA, AA-Briefcase, Agents' Last Exam, and DeepSeek's published agent suite.

• Dated release/confirmation/deadline events and cadence windows inferred from public release dates.

• Additional watched sources for ARC Prize and Agents Last Exam.

• 12 家企业、20 个模型条目，2026-09-10 核实；17 个具有可比的智能与输出价格记录。

• 7 种可切换视图：AA 智力指数 v4.3、Arena 人类偏好、生成速度、SWE-Bench Pro、 Terminal-Bench 4.0（AA 框架）、输出价格、标称上下文。

• 展示原始分值、测量日期、来源链接、初步/厂商自报标签；缺少可比证据时明确显示 N/A。

• 带证据门槛、真实缺口、模型家族色和四种用途自动推荐的五维选型。

• 模型详情按需显示 AA Coding Agent Index、GDPval-AA、AA-Briefcase、 Agents' Last Exam 和 DeepSeek 官方 Agent 套件等专项原始分数。

• 已发布、待确认、截止事件，以及根据公开发布日期推算的历史节奏窗口。

• 额外观察 ARC Prize 和 Agents Last Exam。

Run / 运行

npm start

Open http://127.0.0.1:4174/.

Open V2 at http://127.0.0.1:4174/v2.html.

Open the local radar edition at http://127.0.0.1:4174/radar.html.

Open the model kill line at http://127.0.0.1:4174/radar.html#value-frontier.

Requires Node.js 22 or later. No dependency installation is needed.

需要 Node.js 22 或更新版本，无需安装项目依赖。

Refresh / 刷新

npm run update

The refresh performs one serial pass over a fixed 39-page allowlist. It records freshness, page hashes, failures, and watched terms. It has no login, cookies, proxy, link crawling, endpoint discovery, bypass behavior, or retry loop.

刷新任务每天串行检查固定的 39 个公开页面，记录新鲜度、页面摘要、失败状态和观察词。任务不登录、不带 cookie、不使用代理、不扩展爬取链接、不探测接口、不绕过限制，也不进行循环重试。

Benchmark values stay curated because leaderboards use different harnesses, variants, and update schedules. A source page change becomes a review signal; it does not silently overwrite a score.

Benchmark 分数保留人工确认流程，因为各榜单的 harness、模型变体和更新时间并不统一。来源页面变化会形成复核信号，不会静默覆盖已有分数。

三个页面都展示评测记录的实际日期范围、最新记录距今多少天，以及来源检查成功/失败数。 “来源与变化”展示全部厂商与评测页面，失败项优先，并直接显示错误原因。

更新器保留 lastSuccessAt、lastChangedAt 和最近一次 signalChange（观察词新增/消失及日期）。 失败时保留最近成功的页面摘要和观察词。更换来源 URL 后重新建立基线，首次检查不报新增。 changed 仍然只表示本轮页面摘要发生变化；历史变化日期在后续内容不变时继续保留。 页面摘要可能受动态内容影响，观察词变化也需要人工核实。历史记录从新版检查器运行后逐步积累。

GitHub automation:

.github/workflows/model-radar.yml  # daily source check and snapshot commit
main / repository root             # GitHub Pages branch deployment

GitHub Pages serves the repository root directly from main. This keeps the static deployment independent from the daily source-check workflow.

GitHub Pages 直接发布 main 根目录，静态页面部署与每日来源检查互不依赖。

Verify / 验证

npm test
.\scripts\check.ps1

Reading Scores / 读分规则

Label	Meaning / 含义
INDEPENDENT	Value taken from the named independent leaderboard. / 独立榜单分值。
VENDOR	Vendor-reported result; benchmark and harness version still matter. / 厂商自报结果，仍需核对 benchmark 与 harness 版本。
PRELIM	Preliminary leaderboard row. / 尚未稳定的榜单条目。
LIST	Public list price or advertised context. / 公开标价或标称上下文。
N/A / 榜单未确认	The checked leaderboard has no exact model/version row. / 已核对榜单没有完全同名、同版本条目。
N/A / 未公布同口径成绩	No same-version benchmark and harness result was confirmed. / 未确认同版本、同 benchmark 与同 harness 成绩。
N/A / 公开价格未确认	No directly comparable public API output price was confirmed. / 未确认可直接比较的公开 API 输出价格。

Release timing uses days since latest release / average historical interval. 1.0 means the elapsed time has reached that provider's historical average. It is an attention hint; it carries no launch probability and no official countdown.

历史发布节奏使用“距最近一次发布的天数 / 过去平均发布间隔”。1.0 表示已经走到 该企业过去的平均间隔附近。它只用于安排观察注意力，不代表发布概率或官方倒计时。