# Refined-X 完整公开语料 姓名：Tikoly 简介：Tikoly makes it easy to estimate, purchase and apply the right LLMs to the right agentic jobs. --- # How do I get started with Tikoly? 作者：Tikoly 来源：https://tikoly.com/answers/get-started/ 摘要：Tikoly is currently in private beta. Join the waitlist at tikoly.com to get early access. Beta users get 1M free tokens to test across all supported models. Tikoly is currently in private beta. To get started: 1. **Join the waitlist** at tikoly.com for early access 2. **Get 1M free tokens** to test across all supported models 3. **Explore the Token Dashboard** for real-time cost visibility 4. **Try tikosim CLI** — the free token cost simulator Beta feedback will shape the v1.0 pricing model and feature roadmap. --- # What is Tikoly? 作者：Tikoly 来源：https://tikoly.com/answers/what-is-tikoly/ 摘要：Tikoly is an LLM provider purpose-built for agent makers. It gives you real-time token cost visibility, instant model switching across 15+ providers, and double-entry token accounting with 99.999% auditability. Tikoly is an LLM provider built for agent makers who need real-time cost visibility, multi-model flexibility, and predictable pricing. Unlike traditional API providers that offer opaque billing, Tikoly gives every user a real-time dashboard showing per-request token breakdowns, cumulative cost, and package burn rates. Every token is metered. Every cost is visible. --- # Token Ledger 作者：Tikoly 来源：https://tikoly.com/2027/08/05/builders-token-ledger/ 发布日期：2027-08-05T00:00:00.000Z 更新时间：2027-08-05T00:00:00.000Z 专栏：Article 摘要：A double-entry bookkeeping system providing 99.999% auditability for every token transaction across the Tikoly platform. ## Launch Objective The Token Ledger replaces traditional database accounting with a double-entry engine, providing 99.999% auditability for every token transaction. Every LLM call is debited from the user's account and credited to the provider in a single atomic commit. ## Why It Matters Our accounting engine guarantees consistency even under node failure — a workflow cannot execute unless the ledger confirms a sufficient credit balance. This means agent makers never overdraw, and every token is accounted for with bank-grade precision. Real-time balance checks happen in under 2ms. --- # Token Ledger [中文待翻译] 作者：Tikoly 来源：https://tikoly.com/2027/08/05/builders-token-ledger-zh/ 发布日期：2027-08-05T00:00:00.000Z 更新时间：2027-08-05T00:00:00.000Z 专栏：Article 摘要：Token 账本引擎 [中文待翻译] ## 内容待翻译 此页面英文内容翻译工作正在进行中。 --- # Real-Time Telemetry Pipeline 作者：Tikoly 来源：https://tikoly.com/2027/03/12/builders-real-time-telemetry-pipeline/ 发布日期：2027-03-12T00:00:00.000Z 更新时间：2027-03-12T00:00:00.000Z 专栏：Article 摘要：Sub-10ms real-time token streaming to all Tikoly dashboard users, powered by the RJ Data Center event cluster. ## Launch Objective The Real-Time Telemetry Pipeline brings sub-10ms real-time token streaming to all Tikoly dashboard users. Every token generated at the RJ Data Center is published to our event cluster and rendered on the live dashboard within milliseconds. ## Impact This release eliminates the polling lag that plagues most LLM provider dashboards. Agent makers see token burn update in real time, not on a 30-second refresh cycle. The pipeline also powers the tikosim live display, making the CLI simulator feel responsive even during rapid multi-turn sessions. --- # Real-Time Telemetry Pipeline [中文待翻译] 作者：Tikoly 来源：https://tikoly.com/2027/03/12/builders-real-time-telemetry-pipeline-zh/ 发布日期：2027-03-12T00:00:00.000Z 更新时间：2027-03-12T00:00:00.000Z 专栏：Article 摘要：实时遥测管道 [中文待翻译] ## 内容待翻译 此页面英文内容翻译工作正在进行中。 --- # Tikoly v0.1 Beta 作者：Tikoly 来源：https://tikoly.com/2026/10/18/builders-tikoly-v0-1-beta/ 发布日期：2026-10-18T00:00:00.000Z 更新时间：2026-10-18T00:00:00.000Z 专栏：Article 摘要：Tikoly v0.1 Beta opens the platform to the first cohort of waitlist members with early access to the Token Burn Dashboard, Model Switcher, and tikosim CLI. ## Launch Objective Tikoly v0.1 Beta opens the platform to the first cohort of waitlist members. Early access includes the Token Burn Dashboard, Model Switcher with 5 supported providers, and the tikosim CLI token simulator. ## What's Included Beta users get 1M free tokens to test across all available models. The live dashboard provides real-time telemetry visualization. Feedback from the beta will shape the v1.0 pricing model and feature roadmap. Join the waitlist at tikoly.com to secure your spot. --- # Tikoly v0.1 Beta [中文待翻译] 作者：Tikoly 来源：https://tikoly.com/2026/10/18/builders-tikoly-v0-1-beta-zh/ 发布日期：2026-10-18T00:00:00.000Z 更新时间：2026-10-18T00:00:00.000Z 专栏：Article 摘要：Tikoly v0.1 Beta 内测发布 [中文待翻译] ## 内容待翻译 此页面英文内容翻译工作正在进行中。 --- # Content structure 作者：Tikoly 来源：https://tikoly.com/2026/07/02/content-structure/ 发布日期：2026-07-02T10:00:00.000Z 更新时间：2026-07-02T10:00:00.000Z 专栏：Product 摘要：Tikoly landing content is organized under content/ with articles (explore, news, reports, builders), answers, profile, projects, series, and pages directories. Tikoly's landing content lives under `content/` with an opinionated directory structure: ``` content/ ├── articles/ │ ├── explore/ # Product features │ ├── news/ # Announcements & updates │ ├── reports/ # Research deep-dives │ └── builders/ # Product milestones ├── answers/ # Curated Q&A ├── pages/ # Static pages (about, friends) ├── profile/ # Company identity & resume ├── projects/ # Product showcase └── series/ # Article series ``` Each section has optional `zh/` subdirectories for Chinese translations. Machine-readable endpoints (`/llms.txt`, `/openapi.json`, `/.well-known/`) are generated at build time from this corpus. --- # Welcome to Tikoly 作者：Tikoly 来源：https://tikoly.com/2026/07/01/welcome-to-tikoly/ 发布日期：2026-07-01T10:00:00.000Z 更新时间：2026-07-01T10:00:00.000Z 专栏：Research 摘要：Tikoly is the LLM provider built for agent makers. This site documents features, research reports, product news, and builder events. Welcome to Tikoly — the LLM provider built for agent makers. ## What you'll find here - **Explore** — Product features: Token Burn Dashboard, Model Switcher, Cost Simulator - **Reports** — Deep-dives into LLM economics, token optimization, and agent architecture - **News** — Product updates, platform announcements, and engineering briefings - **Builders** — Product launches, milestones, and community events ## Agent-friendly This site exposes machine-readable surfaces for AI agents: `/llms.txt` for content discovery, OpenAPI for programmatic access, and MCP endpoints for live querying. Ask questions naturally via the Ask bar. --- # Cost Simulator 作者：Tikoly 来源：https://tikoly.com/2026/07/01/explore-cost-simulator/ 发布日期：2026-07-01T00:00:00.000Z 更新时间：2026-07-01T00:00:00.000Z 专栏：Article 摘要：Pre-purchase token cost simulation — run real task profiles against any model and see exact dollar burn before committing to a package. ## Mission Objective Cost Simulator wraps pi.dev's agent harness in a pricing engine that meters token burn in real time. Users pick a task profile (code-gen, translate, analysis, chat-sim, thinking), select a model, and watch cumulative cost accrue against tikoly.com package tiers. ## Current Operations The pricing interceptor subscribes to the agent event stream and overrides provider list prices with tikoly.com's rate card. Session files produce JSONL traces with per-request cost breakdowns. Users can fork sessions to compare the same task across different models side by side. --- # 成本模拟器 作者：Tikoly 来源：https://tikoly.com/2026/07/01/explore-cost-simulator-zh/ 发布日期：2026-07-01T00:00:00.000Z 更新时间：2026-07-01T00:00:00.000Z 专栏：Article 摘要：购买前的 Token 成本模拟——针对任何模型运行真实任务配置文件，在承诺购买套餐前查看精确的美元消耗。 ## 任务目标 成本模拟器将 pi.dev 的智能体框架包装在定价引擎中，实时计量 Token 消耗。用户选择任务配置文件（代码生成、翻译、分析、聊天模拟、深度思考），选择模型，并观察累计成本针对 tikoly.com 套餐级别的增长。 ## 当前状态 定价拦截器订阅智能体事件流，用 tikoly.com 的费率表覆盖提供商标价。会话文件生成包含每次请求成本明细的 JSONL 追踪记录。用户可分支会话，并排比较同一任务在不同模型上的成本。 --- # Model Switcher 作者：Tikoly 来源：https://tikoly.com/2026/07/01/explore-model-switcher/ 发布日期：2026-07-01T00:00:00.000Z 更新时间：2026-07-01T00:00:00.000Z 专栏：Article 摘要：Instantly switch between 15+ LLM models mid-session and see the cost impact update in real time — DeepSeek, OpenAI, Anthropic, Google, Mistral and more. ## Mission Objective Model Switcher lets agent makers cycle through models with Ctrl+P, open a fuzzy-search overlay with Ctrl+L, or set scoped model lists via --models flag. Thinking levels (off → xhigh) adjust reasoning token budgets on the fly. ## Current Operations The router dispatches requests through pi-ai's unified multi-provider layer. Each model switch updates the footer bar with new token/cost projections. The extension API allows custom rate cards to override provider list prices for tikoly.com pricing. --- # 模型切换器 作者：Tikoly 来源：https://tikoly.com/2026/07/01/explore-model-switcher-zh/ 发布日期：2026-07-01T00:00:00.000Z 更新时间：2026-07-01T00:00:00.000Z 专栏：Article 摘要：在会话中即时切换 15+ 种 LLM 模型，实时查看成本变化——支持 DeepSeek、OpenAI、Anthropic、Google、Mistral 等。 ## 任务目标 模型切换器让智能体构建者可通过 Ctrl+P 循环切换模型，通过 Ctrl+L 打开模糊搜索覆盖层，或通过 --models 标志设置作用域模型列表。思考级别（关闭→极高）可实时调整推理 Token 预算。 ## 当前状态 路由器通过 pi-ai 的统一多提供商层分发请求。每次模型切换都会更新页脚栏中的新 Token/成本预估。扩展 API 允许自定义费率表覆盖提供商标价以应用 tikoly.com 定价。 --- # Token Burn Dashboard 作者：Tikoly 来源：https://tikoly.com/2026/07/01/explore-token-burn-dashboard/ 发布日期：2026-07-01T00:00:00.000Z 更新时间：2026-07-01T00:00:00.000Z 专栏：Article 摘要：Real-time token usage monitoring across all LLM calls — input, output, cached, and reasoning tokens displayed per-request and cumulative. ## Mission Objective Token Burn Dashboard gives agent makers full visibility into how every LLM call consumes tokens. Input, output, cached reads, and reasoning tokens are all tracked per-request and rolled up into session-level totals. ## Current Operations The dashboard renders live telemetry from the event stream, updating cost and token counts on every `message_end` event. Users can filter by model, time range, or task profile. Package burn gauges show remaining quota against Starter, Pro, and Enterprise tiers. --- # Token 消耗仪表盘 作者：Tikoly 来源：https://tikoly.com/2026/07/01/explore-token-burn-dashboard-zh/ 发布日期：2026-07-01T00:00:00.000Z 更新时间：2026-07-01T00:00:00.000Z 专栏：Article 摘要：实时监控所有 LLM 调用的 Token 使用情况——输入、输出、缓存和推理 Token，按请求和累计显示。 ## 任务目标 Token 消耗仪表盘为智能体构建者提供每次 LLM 调用的完整可见性。输入、输出、缓存读取和推理 Token 均按请求追踪，并汇总为会话级总计。 ## 当前状态 仪表盘通过实时事件流渲染遥测数据，在每次 `message_end` 事件时更新 Token 和成本计数。用户可按模型、时间范围或任务配置文件筛选。套餐消耗条显示 Starter、Pro 和 Enterprise 级别的剩余配额。 --- # Tikoly Opens Waitlist for Agent-First LLM Provider 作者：Tikoly 来源：https://tikoly.com/2026/04/18/news-tikoly-opens-waitlist-for-agent-first-llm-provider/ 发布日期：2026-04-18T00:00:00.000Z 更新时间：2026-04-18T00:00:00.000Z 专栏：Article 摘要：Agent makers can now register for early access to Tikoly — the LLM provider built for multi-turn agent workflows with real-time cost visibility and 99.999% auditability. ## Waitlist Now Open Tikoly officially opened its waitlist today, giving agent makers early access to a purpose-built LLM provider that treats token consumption as a first-class metric. Unlike traditional API providers that offer opaque billing, Tikoly gives every user a real-time dashboard showing per-request token breakdowns, cumulative cost, and package burn rates. ## What Early Access Includes Waitlist members will receive priority access to the Tikoly v0.1 Beta, including the Token Burn Dashboard, Model Switcher with 15+ supported models, and the tikosim CLI token simulator. Each beta user gets 1M free tokens to test across all available providers. ## Built for Agent Makers Tikoly's architecture — real-time telemetry, durable workflow orchestration, and double-entry accounting — was designed from the ground up for agent workloads that require reliable execution, real-time monitoring, and predictable pricing. No more surprise bills at the end of the month. --- # Tikoly 开启面向智能体构建者的 LLM 提供商内测申请 作者：Tikoly 来源：https://tikoly.com/2026/04/18/news-tikoly-opens-waitlist-for-agent-first-llm-provider-zh/ 发布日期：2026-04-18T00:00:00.000Z 更新时间：2026-04-18T00:00:00.000Z 专栏：Article 摘要：智能体构建者现可注册 Tikoly 早期访问——专为多轮智能体工作流构建的 LLM 提供商，具备实时成本可见性和 99.999% 可审计性。 ## 内测申请现已开放 Tikoly 今日正式开放内测申请，让智能体构建者能够提前使用一款将 Token 消耗作为一级指标的专用 LLM 提供商。与提供不透明账单的传统 API 提供商不同，Tikoly 为每位用户提供实时仪表盘，显示每次请求的 Token 明细、累计成本和套餐消耗率。 ## 早期访问包含内容 内测成员将优先获得 Tikoly v0.1 Beta 的访问权限，包括 Token 消耗仪表盘、支持 15+ 种模型的模型切换器以及 tikosim CLI Token 模拟器。每位 Beta 用户将获得 100 万免费 Token，可在所有支持的提供商上测试。 ## 为智能体构建者而生 Tikoly 的架构——实时遥测、持久工作流编排和复式记账——专为需要可靠执行、实时监控和可预测定价的智能体工作负载而设计。月末不再有意外账单。 --- # COST / DEEPSEEK V4 ANALYSIS 作者：Tikoly 来源：https://tikoly.com/2026/04/07/report-cost-deepseek-v4-analysis/ 发布日期：2026-04-07T00:00:00.000Z 更新时间：2026-04-07T00:00:00.000Z 专栏：Article 摘要：Per-million-token cost breakdown across all Tikoly-supported models. DeepSeek v4 Flash delivers the lowest cost per token for high-throughput agent tasks. ## Summary Tikoly's pricing team analyzed per-million-token costs across 15 supported models, factoring input, output, cached input, and reasoning token rates. DeepSeek v4 Flash emerges as the most cost-effective option for high-throughput agent workloads, while premium reasoning models (o1, DeepSeek R1) command higher rates for complex multi-step tasks. ## Key Findings - **Input token range**: $0.15/M (DeepSeek v4 Flash) to $15.00/M (o1 with reasoning) - **Cached input discount**: 50% off input rate on all supported models when cache hits occur - **Reasoning token premium**: Thinking models can consume 5-10x more tokens on internal reasoning than visible output - **Package impact**: The Enterprise tier (100M tokens at $30) brings effective per-token cost below $0.30/M for most models ## Methodology Each model was benchmarked against five standard task profiles: code generation, translation, PR analysis, multi-turn chat, and deep research. Token counts were captured from pi-ai's Usage interface and multiplied against tikoly.com's rate card. The spectrum above shows relative cost intensity across model tiers. --- # 成本 / DEEPSEEK V4 分析 作者：Tikoly 来源：https://tikoly.com/2026/04/07/report-cost-deepseek-v4-analysis-zh/ 发布日期：2026-04-07T00:00:00.000Z 更新时间：2026-04-07T00:00:00.000Z 专栏：Article 摘要：Tikoly 支持的所有模型的每百万 Token 成本分析。DeepSeek v4 Flash 为高吞吐智能体任务提供最低的每 Token 成本。 ## 摘要 Tikoly 的定价团队分析了 15 个支持模型的每百万 Token 成本，综合考虑了输入、输出、缓存输入和推理 Token 费率。DeepSeek v4 Flash 是高吞吐智能体工作负载中最具成本效益的选择，而高级推理模型（o1、DeepSeek R1）因其复杂多步骤任务能力而定价较高。 ## 主要发现 - **输入 Token 范围**：$0.15/百万（DeepSeek v4 Flash）至 $15.00/百万（o1 含推理） - **缓存输入折扣**：所有支持模型的缓存命中均享受输入费率五折优惠 - **推理 Token 溢价**：思考模型的内部推理 Token 消耗可达可见输出的 5-10 倍 - **套餐影响**：Enterprise 级别（1 亿 Token，$30）将大多数模型的有效每 Token 成本降至 $0.30/百万以下 ## 方法论 每个模型针对五个标准任务配置文件进行了基准测试：代码生成、翻译、PR 分析、多轮对话和深度研究。Token 计数通过 pi-ai 的 Usage 接口捕获，并与 tikoly.com 的费率表相乘计算。 --- # tikosim Token Simulator Now Available as Free CLI 作者：Tikoly 来源：https://tikoly.com/2026/03/30/news-tikosim-token-simulator-now-available-as-free-cli/ 发布日期：2026-03-30T00:00:00.000Z 更新时间：2026-03-30T00:00:00.000Z 专栏：Article 摘要：Download tikosim — a real-time LLM token cost simulator that lets agent makers compare models, task profiles, and pricing tiers before buying a token package. ## Free CLI Tool Released Tikosim, the real-time LLM token cost simulator, is now available as a free download from tikoly.com. Built on top of pi.dev's agent harness, tikosim lets users run real task profiles — code generation, translation, PR analysis, multi-turn chat, and deep research — against any supported model and watch token burn tick in real time. ## How It Works Users pick a task profile, select a model via Ctrl+P, and tikosim executes the task through pi-ai's multi-provider layer. A pricing interceptor applies tikoly.com's rate card and displays per-request token breakdowns (input, output, cached, reasoning) alongside cumulative cost and package burn gauges. ## Why It Matters LLM token pricing is abstract. Customers see "X tokens per dollar" but have no visceral sense of how that maps to real work. Tikosim closes that gap by showing exactly how many code reviews 100K tokens buys, or how fast a reasoning session burns through a premium package. Try it at tikoly.com/tikosim. --- # tikosim Token 模拟器现以免费 CLI 形式发布 作者：Tikoly 来源：https://tikoly.com/2026/03/30/news-tikosim-token-simulator-now-available-as-free-cli-zh/ 发布日期：2026-03-30T00:00:00.000Z 更新时间：2026-03-30T00:00:00.000Z 专栏：Article 摘要：下载 tikosim——一款实时 LLM Token 成本模拟器，让智能体构建者在购买 Token 套餐前比较模型、任务配置文件和定价级别。 ## 免费 CLI 工具发布 Tikosim，实时 LLM Token 成本模拟器，现可从 tikoly.com 免费下载。基于 pi.dev 的智能体框架构建，tikosim 让用户针对任何支持的模型运行真实任务配置文件——代码生成、翻译、PR 分析、多轮对话和深度研究——并实时观察 Token 消耗。 ## 工作原理 用户选择任务配置文件，通过 Ctrl+P 选择模型，tikosim 通过 pi-ai 的多提供商层执行任务。定价拦截器应用 tikoly.com 的费率表，并显示每次请求的 Token 明细（输入、输出、缓存、推理）以及累计成本和套餐消耗条。 ## 为何重要 LLM Token 定价是抽象的。用户看到"每美元 X Token"，但对其如何映射到实际工作缺乏直观感受。Tikosim 通过精确展示 10 万 Token 能购买多少代码审查，或一次推理会话能以多快速度烧完高级套餐，弥补了这一差距。 --- # USAGE+TOKEN BURN PATTERNS 作者：Tikoly 来源：https://tikoly.com/2026/03/16/report-usage-token-burn-patterns/ 发布日期：2026-03-16T00:00:00.000Z 更新时间：2026-03-16T00:00:00.000Z 专栏：Article 摘要：Analysis of real-world token consumption across 10,000 agent sessions reveals that context accumulation is the dominant cost driver for multi-turn agent workflows. ## Summary Analysis of 10,000 anonymized agent sessions on the tikosim simulator reveals that context accumulation across turns is the single largest cost driver. A 20-turn chat session costs 5.2x more than 20 individual single-turn requests due to growing system context with each turn. ## Key Patterns - **Context bloat**: Average session context grows 340% from turn 1 to turn 20 - **Cached input efficiency**: Providers with cache-read discounts save 40-60% on repeated system prompts - **Reasoning token variance**: 'thinking' profile sessions show 8.3x standard deviation in token burn compared to 'code-gen' - **Optimal model switching**: Switching from a premium reasoning model to a flash model after the first planning turn reduces total cost by 62% ## Recommendations Agent makers should set per-session token budgets, leverage cached input where available, and consider hybrid model strategies that use cheap models for routine turns and expensive reasoning models only when needed. --- # TOKEN 消耗模式 作者：Tikoly 来源：https://tikoly.com/2026/03/16/report-usage-token-burn-patterns-zh/ 发布日期：2026-03-16T00:00:00.000Z 更新时间：2026-03-16T00:00:00.000Z 专栏：Article 摘要：对 10,000 个智能体会话的 Token 消耗分析显示，上下文累积是多轮智能体工作流的主要成本驱动因素。 ## 摘要 对 tikosim 模拟器上 10,000 个匿名智能体会话的分析显示，跨轮次的上下文累积是最大的单一成本驱动因素。由于每轮系统上下文不断增长，20 轮对话会话的成本是 20 个独立单轮请求的 5.2 倍。 ## 关键模式 - **上下文膨胀**：从第 1 轮到第 20 轮，平均会话上下文增长 340% - **缓存输入效率**：提供缓存读取折扣的提供商可在重复系统提示上节省 40-60% - **推理 Token 差异**：'深度思考'配置文件的 Token 消耗标准差是'代码生成'的 8.3 倍 - **最优模型切换**：在首轮规划后从高级推理模型切换到快速模型，可降低 62% 的总成本 ## 建议 智能体构建者应设置每会话 Token 预算，尽可能利用缓存输入，并考虑混合模型策略——常规轮次使用廉价模型，仅在需要时使用昂贵推理模型。 --- # Tikoly Partners with RJ Data Center for Sub-10ms Inference 作者：Tikoly 来源：https://tikoly.com/2026/02/11/news-tikoly-partners-with-rj-data-center-for-sub-10ms-inference/ 发布日期：2026-02-11T00:00:00.000Z 更新时间：2026-02-11T00:00:00.000Z 专栏：Article 摘要：Direct peering with RJ Data Center brings Tikoly's median inference latency below 10ms, enabling real-time agent applications that cannot tolerate jitter. ## RJ Data Center Partnership Tikoly has established a direct peering partnership with RJ Data Center to host GPU inference nodes adjacent to the telemetry cluster. The arrangement eliminates intermediary hops and brings median time-to-first-token below 340ms across all supported models. ## Infrastructure Details The RJ Data Center deployment runs dedicated inference nodes for DeepSeek, OpenAI, Anthropic, Google, and Mistral models. Workflows orchestrate multi-step agent tasks with durable execution guarantees. Our double-entry ledger provides accounting for every token generated. ## What This Means for Agent Makers Sub-10ms telemetry means the dashboard updates in real time — no polling, no refresh buttons. Agent makers see token burn, cost accrual, and package depletion as it happens. This level of visibility is critical for teams that need to manage costs across hundreds of concurrent agent sessions. --- # Tikoly 与 RJ 数据中心达成合作，实现亚 10ms 推理 作者：Tikoly 来源：https://tikoly.com/2026/02/11/news-tikoly-partners-with-rj-data-center-for-sub-10ms-inference-zh/ 发布日期：2026-02-11T00:00:00.000Z 更新时间：2026-02-11T00:00:00.000Z 专栏：Article 摘要：与 RJ 数据中心的直接对等连接使 Tikoly 的中位推理延迟低于 10ms，为无法容忍抖动的实时智能体应用提供支持。 ## RJ 数据中心合作 Tikoly 已与 RJ 数据中心建立直接对等合作，在遥测集群旁托管 GPU 推理节点。这一安排消除了中间跳转，使所有支持模型的中位首 Token 时间降至 340ms 以下。 ## 基础设施详情 RJ 数据中心部署运行 DeepSeek、OpenAI、Anthropic、Google 和 Mistral 模型的专用推理节点。工作流编排多步骤智能体任务，具备持久执行保证。我们的复式账本为每 Token 提供记账。 ## 这对智能体构建者的意义 亚 10ms 遥测意味着仪表盘实时更新——无需轮询，无需刷新按钮。智能体构建者可以实时看到 Token 消耗、成本累积和套餐消耗。这种级别的可见性对于需要管理数百个并发智能体会话成本的团队至关重要。 --- # PERFORMANCE / LATENCY BENCHMARKS 作者：Tikoly 来源：https://tikoly.com/2026/02/02/report-performance-latency-benchmarks/ 发布日期：2026-02-02T00:00:00.000Z 更新时间：2026-02-02T00:00:00.000Z 专栏：Article 摘要：End-to-end latency measurements across Tikoly's provider network. The telemetry pipeline delivers sub-10ms median latency for token streaming. ## Summary Tikoly's telemetry backbone delivers sub-10ms median latency between the RJ Data Center and end-user dashboards. The 99th percentile remains under 42ms even under maximum load, making Tikoly suitable for real-time agent applications that cannot tolerate jitter. ## Benchmark Results - **Median TTFB** (time to first token): 340ms across all providers - **P99 TTFB**: 890ms for flash models, 2.4s for reasoning models - **Telemetry pipeline latency**: 3.2ms median (publish to dashboard update) - **Accounting commit**: 1.8ms median per debit/credit transaction ## Infrastructure The RJ Data Center runs dedicated inference nodes. Workflows orchestrate multi-step agent tasks with durable execution guarantees. Our double-entry ledger accounts for every token generated — 99.999% auditability on all transactions. --- # 性能 / 延迟基准测试 作者：Tikoly 来源：https://tikoly.com/2026/02/02/report-performance-latency-benchmarks-zh/ 发布日期：2026-02-02T00:00:00.000Z 更新时间：2026-02-02T00:00:00.000Z 专栏：Article 摘要：跨 Tikoly 提供商网络的端到端延迟测量。遥测管道为 Token 流提供低于 10ms 的中位延迟。 ## 摘要 Tikoly 的遥测骨干网络在 RJ 数据中心与用户仪表盘之间提供低于 10ms 的中位延迟。即使在最大负载下，99 百分位仍保持在 42ms 以下，使 Tikoly 适用于无法容忍抖动的实时智能体应用。 ## 基准测试结果 - **中位 TTFB**（首 Token 时间）：所有提供商平均 340ms - **P99 TTFB**：快速模型 890ms，推理模型 2.4s - **遥测管道延迟**：3.2ms 中位（发布到仪表盘更新） - **记账提交**：每笔借记/贷记交易 1.8ms 中位 ## 基础设施 RJ 数据中心运行专用推理节点。工作流编排多步骤智能体任务，具备持久执行保证。我们的复式账本为每 Token 提供记账——所有交易 99.999% 可审计。