把你的想法，变成可以交付的产物。
A workstation that turns intent into artifact.
在同一个界面里完成所有事：说出你要什么、让工具与模型协作、观察每一步的执行、留下可以回溯的产物。底层是子午线协议——38 个工具、5 段级联、每一次运行都可以复现。
One surface where intent lands, tools orchestrate, models compete, a judge votes, artifacts ship, and versions archive — no window‑switching, no lost context. Powered by the Meridian protocol: 38 tools, 5‑stage cascade, end‑to‑end auditable.
聊天工具给你的，是一段凭感觉写出来的文字。Meridian 做的是另一件事：把你说的话拆成可执行的步骤，派给 38 个工具去跑，让多个模型并行，再交给三位独立裁判评审，最后落成一份带版本的产物。同样的输入，明天跑、换模型跑——结果一致。这不是运气，是工程。
Chatbots guess. Meridian decomposes your intent into an executable plan, dispatches it across 38 tools, pipes the result through judge voting, and lands a versioned artifact. Run it tomorrow, run it on a new model — same artifact. This is engineering, not luck.
工作站的四面墙——意图从这里进去，产物从这里出来，过程在这里看得见，历史在这里留得下。
The four sides of the workstation: intent enters, artifacts ship, process is observed, history is kept.
Serpentine 是完整级联：头部先拆分意图，提示词拼装上下文，复核做独立交叉检验，主体并行跑旗舰模型并产出媒体，尾部交由三位独立裁判投票并注入引用、写入版本。任意一关未通过，整条管线立即回退。
Serpentine is the full cascade: head decomposes, prompt assembles, review cross‑checks, body races flagships with judge voting, tail closes with citations and version pin. Any gate can halt and roll back.
两队模型。大脑负责编排与推理，蛇纹石的每一关挂载不同档位。工具 AI 负责生成图像、视频、语音、代码，由主体阶段按需调用。列表随上游发布持续更新。
Two teams. The brain models run orchestration and reasoning — one tier per Serpentine gate. Tool models generate images, video, audio, code — invoked by body on demand. Registry updates continuously.
蛇纹石五关——头部、提示词、复核、主体、尾部——分别挂载不同档位的旗舰推理模型。高风险关口使用 high 档，常规使用 medium 档。
Across the five Serpentine gates, each mounts a flagship reasoning model at a tier. Critical gates run at high; routine ones at medium.
由主体阶段调用。经 fal.ai 统一接口路由，文生图与图生图(edit / img2img)在后端注册表中一一对应，可替换、可拓展。
Invoked by the body stage. Routed via fal's unified interface; text‑to‑image and edit/img2img endpoints map 1‑to‑1 in the backend registry — swappable and extensible.
在我们的编排体系里，深度研究不是一次长上下文的生成，而是一组可重复的实验。一个问题被提出后，Meridian 并不会立刻交给单一模型作答；它先走头部阶段，把问题拆成若干子命题，为每个子命题拟定证据要求，再把这些检索与推理任务派发给不同档位的大脑模型去并行执行1。
这里登场的是三个各自擅长不同工作的旗舰：Gemini Deep Research 负责横向扫描与资料归集，它最强的是把海量来源折叠成一份结构化摘要；GPT 5.4 Pro XHigh 负责把摘要展开成严谨的中文论证链条，强点在于长逻辑链下的句法稳定；Claude Opus 4.7 adaptive 则承担审校与反驳任务，它能在阅读自己刚刚生成的结论时做自我批评，补齐证据或标出漏洞。这三者在主体阶段并行，在尾部阶段互为裁判，三票一致才写入产物2。
In our orchestration system, deep research is not a single long‑context generation; it is a set of reproducible experiments. When a question arrives, Meridian does not forward it to a single model. The head stage first decomposes it into sub‑claims, declares the evidence each sub‑claim requires, and dispatches retrieval and reasoning tasks in parallel across brain models of different tiers1.
Three flagships are on stage. Gemini Deep Research handles lateral scanning and corpus assembly — its strength is folding large volumes of sources into a structured digest. GPT 5.4 Pro XHigh expands the digest into rigorous argument chains, stable across long logical dependencies. Claude Opus 4.7 adaptive runs audit and rebuttal, reading its own output with self‑criticism, supplementing evidence or marking gaps. The three run in parallel during the body stage and vote on each other at the tail; only unanimous approval ships as an artifact2.
这种三路合作的结构并不是为了产出更长的文字。它的目的是把认识论上的冗余引入单一问题的处理过程。一个研究结论若想被企业采用，仅仅"语言通顺"是不够的——它需要在事实、推理、反例三件事上同时站得住。Meridian 的深研管线用三位模型分别承担这三件事，而不是指望一个模型同时具备三种不同的思维习惯3。
交付产物是一份带引用、带版本、可审计的artifact。它不是一段文字，是一个可被 diff、可被回溯、可被 Git 追踪的对象。每一处引用都能点开原始来源；每一条推论都能展开它的支撑证据；每一次裁判投票都有文本记录。这意味着如果你下周再跑一遍同样的追问，你会得到一份字符级别可比较的新产物——差异只可能来自上游数据的更新，不来自模型情绪的起伏4。
This tripartite structure is not about producing more text. Its purpose is to inject epistemic redundancy into the handling of a single question. For a research conclusion to be usable by an enterprise, linguistic fluency is not enough — it must simultaneously hold across evidence, reasoning, and counter‑example. Meridian's deep research pipeline assigns those three concerns to three models, rather than hoping one model possesses all three cognitive habits at once3.
The deliverable is a cited, versioned, auditable artifact. Not prose; a diffable, retrievable, Git‑trackable object. Every citation opens its source. Every inference expands its supporting evidence. Every judge vote has a text record. Re‑running the same question next week yields character‑level comparable output — any difference comes from upstream data changes, not from model mood drift4.
Meridian 象限仪不是四个面板，是四个互相制衡的角色——思想、行动、编排、记忆。每个象限拿到属于它的那一份信息，做属于它的那一份决定；没有任何一个象限同时握住全部上下文与全部权限。这是我们的协作范式，也是我们愿意把虚拟机引入工作区的前提。
Meridian's Quadrants are not four panels; they are four roles held in tension — thought, action, topology, memory. Each quadrant receives the slice of information it needs and makes the decision it owns. No single quadrant simultaneously holds full context and full authority. This is our cooperation paradigm — and the precondition for introducing a virtual machine into the workspace.
没有任何一个象限同时拥有全部信息和全部权限。思想 + 行动 分离 权限 + 信息 分离 监督 + 执行 分离
No quadrant holds full information and full authority at the same time.Thought ÷ Action Authority ÷ Information Oversight ÷ Execution
每次任务完成后,Q3 的派发脉络、Q1 的推理过程、Q2 的操作录像,在脱敏之后全部归档进 Q4。下一次 Q3 做编排决策时,它收到的不只是当前任务描述——还有 Q4 基于历史归档提出的模式建议:重复的问题、被反复标记的坑、针对这个用户的最佳实践。
这个闭环不依赖模型记得什么。它依赖一个独立的、结构化的、可审计的知识层。模型可以换，Q4 留下的经验不换。未来它也会对接具身硬件——当手臂抓握一个玻璃杯失败三次，Q4 就会把这件事写成 Q3 下次的前置条件。
After each task, Q3's dispatch flow, Q1's reasoning trace, and Q2's operation recording are sanitized and archived into Q4. Next time Q3 makes an orchestration decision, it receives not only the current task description — but also pattern advice from Q4 based on historical archives: recurring problems, repeatedly flagged pitfalls, user‑specific best practices.
The loop does not rely on any model remembering. It relies on an independent, structured, auditable knowledge layer. The models can change; what Q4 has learned does not. In the future, it will also interface with embodied hardware — when an arm fails three times to grip a glass, Q4 will encode that into Q3's preconditions for next time.
Q2 能调用带密钥的外部服务——Vercel 部署、S3 上传、支付接口——却看不到密钥明文。做法借了餐厅后厨的隐喻：厨师（Q2 的 sub-agent）从出餐口递出订单，前台（密钥代理）接单后用真实凭证完成调用，再把结果递回后厨。AI 拿到的是结果，不是凭证。
沙盒虚拟机内部还有一个议会 · Parliament：多个 sub-agent 互相审查彼此的中间产物——一个建环境、一个拉文件、一个写代码、一个负责回滚检查。谁都能在下一步被其他人否决。
Q2 can invoke credentialed external services — Vercel deploys, S3 uploads, payment APIs — without ever seeing the key plaintext. The metaphor is a restaurant pass‑through: the cook (Q2 sub‑agent) pushes the order through the service window; the front desk (key broker) completes the call with real credentials and pushes back the result. The AI receives the result, never the credentials.
Inside the sandboxed VM there is also a Parliament: multiple sub‑agents review each other's intermediate outputs — one provisions the environment, one pulls files, one writes code, one checks rollback. Any next step can be vetoed by another.
Q2 当前是多模态画布形态：文件、图像、代码、artifact 四视图并存，支持生成、编辑、审校、导出。暂无独立虚拟机，所有操作在我们自己的服务端内收敛执行。
Q2 today is a multimodal canvas: files, images, code, and artifacts coexist with generate / edit / review / export. No separate VM yet — all operations are converged inside our controlled server plane.
Q2 升级为沙盒化虚拟机，里面跑多个互审 sub-agent，通过出餐口协议对外调用带密钥的服务。移动端可访问、可托付长时任务，下班回家之后仍在跑。
Q2 upgrades to a sandboxed VM housing multiple peer‑reviewing sub‑agents. Outbound calls go through the Service Window protocol; credentials never enter the AI. Mobile‑accessible; long tasks continue after you log off.
六种外壳，一键切换，颜色会跟着你的一天走。点下方的色块试试——顶部 hero 会实时换色，光晕、进度条、品牌色也一并联动。
Six shells · one click · colors that follow your day. Click a swatch below — the hero above swaps live, along with the spotlight, progress bar, and brand accent.
过去一年每一起 P0 故障，最终回复客户的都是写那行代码的工程师本人。团队背景均为计算机科学、系统工程、分布式基础设施方向，arxiv 是日常读物。
Every P0 of the past year — you ended up talking to the engineer who wrote the code. Backgrounds: CS, systems, distributed infra. arxiv is daily reading.
38 个工具只是起点。我们为客户开发过内部迁移命令行、账单对账管线、跨仓库 release note 合并脚本——这些都是后期按需加上的。把场景讲清楚：能做就做，做不了会直接告知。
38 tools is the floor. We've shipped custom migration CLIs, billing reconciliation pipelines, cross‑repo release‑note merge. Describe the job — we'll tell you straight if we can do it.
从 OpenAI、Anthropic 或本地 vLLM 栈迁移过来：我们先阅读你现有代码，再提出适配方案，最后自己完整跑一遍迁移。不是丢一份文档让客户自行复制粘贴。
Moving off OpenAI / Anthropic / vLLM: we read your current code, draft the adapter, run the migration ourselves. Not a doc we ask you to copy‑paste.
产物结果异常时，内部复核通道最快一个工作日接手处理。这不是写在 SLA 条款里的承诺——我们团队自身也每天在使用这套运行时。
If an output looks wrong, the internal review channel picks it up within one business day. Not an SLA clause — we run on the same runtime.
套餐只代表积分额度的多少。不限用量、不限工具、不限级联段数——所有模型、所有流程、Ultra 五段级联全部档位都可使用。成本按每次调用的实际用量动态扣除,用多少算多少。
Plans are just credit bundles. No cap on usage, tools, or cascade depth — every model, every stage, Ultra included. Credits are deducted dynamically based on actual usage.
* 积分按每次调用的实际成本动态扣除。不同模型、不同档位扣除额度不同，不设固定费率表。详细明细以账单为准。* Credits are debited dynamically against real per‑call cost. Different models at different tiers consume different amounts — no fixed table. Exact accounting on your invoice.
B2B 专供。入门档 ¥6,980 / 月，开放全部模型、全部工具、完整级联。MAX 20X 与私有化部署按客户当前技术栈另行配置，请与销售对接。
B2B only. Entry ¥6,980/mo — all models, all tools, full cascade included. MAX 20X and on‑prem deployments are configured per your existing stack; contact sales to proceed.
一台把想法做成产物的 AI 工作站。节点部署于 Sydney、Singapore、Hong Kong。
An AI workstation that turns intent into artifact. Nodes: Sydney · Singapore · Hong Kong.