不可变版本 / Immutable version 2026-10-25.3. 本副本不加载分析或应用脚本 / No analytics or app scripts.
分别约定自有机器模型转发与按时租用 GPU 机器；两者的费用和数据责任不同。
2026-10-25.3 · 2026-10-25
本附件适用于 Laintas Compute 的相应功能，与平台服务条款、隐私政策、数据用途说明及收费与退款政策共同适用。仅在本附件明确的产品事项上补充或优先于总条款；强制性法律及双方有效的个别协商约定优先。以购买或使用时明确提供且依法有效接受的版本为准，网页更新不自动改变已有订单。
模型在您的设备或自选运行时执行，网关转发请求与回复，并统计转发字节、时间、token 数和状态等元数据。该路径不按 token 收 Laintas 推理费，流量额度按会员或加购规则计算；耗尽、离线或额度不可读时拒绝或暂停，不静默改用付费模型。设备离线不等于凭据撤销；删除设备并移除本机凭据可终止后续接入。电力、硬件、网络和您自选运行时费用由相应合同决定。
租机使用账户余额，不包含在会员推理额度内。按下单显示的 GPU、每小时价格、时长或不限结束时间方式提供。当前实现以服务成功连入机器并标记运行的 started_at 开始计时，而不是仅以下单时间计费；部署模型时，之后的下载和加载时间属于占用时间。通常每十分钟结算一次，按实际经过时间累计，最终按分结清，不等于每开始十分钟必收完整十分钟；页面如另列固定费用须同时确认。停止请求后不再因销毁重试额外延长用户计费截止时间。
未成功进入计时运行状态的分配失败或启动超时，不按正常运行时间收费。机器已就绪后，您选择的模型不兼容、下载或加载失败可能仍占用租机时间，可停止租用并按原因请求核查；不因模型失败就保证全部租机时间免费。因 Laintas 可归责故障造成未交付、重复计时或错误收费的，应更正并依法补救；上游故障按故障时间、记录、订单和各方义务处理，不仅因是第三方就排除退款。
固定租期届满、主动收回、余额不足、安全处置或上游中断可能使机器停止并销毁。不限结束时间不代表无限信用，余额不足下一结算区间时可能自动收回。机器和磁盘是临时环境，收回后不保留数据；即使在租期内也应备份，不把服务当作唯一存储。导出需要在终止前完成，但此风险提示不免除平台因自身过错依法应承担的责任。SSH 私钥应妥善保存且不得公开。
租用机器可由 Lium、Vast.ai 等当次实际启用的计算基础设施提供，地区以所选机器及订单可得信息为准；上游及机主可能控制底层环境。您须有权接入设备和运行模型，遵守模型权重、运行时、下载站点和基础设施的许可及用途限制，不能因可下载就推定可转售。安装器可按您选择安装引擎、下载模型或配置开机启动，应核对下载来源和机器权限。私有模型访问令牌只应用于授权用途。
当前租机在 SSH 服务成功连接并进入运行状态后起算，之后下载模型、装载权重和准备推理仍占用机器并计时；不是首次成功回答才开始收费。账务通常每十分钟结算已发生时长，并在结束时按实际占用时间补结、依账本精度舍入到分，不是每启动一次就收十分钟最低费。停止请求成功记录后，以 stop_requested_at 或相应结束时间作为用户费用截点；销毁重试本身不应把截点后的时间继续算给用户。遇到状态不一致，可凭租用号请求核对起算、截点和扣款。
GPU 型号、显存、磁盘、地区、可用性与小时价应在租用确认时核对，列表估价不是任意期间固定可用的容量承诺。部署前就失败且从未进入正常计费状态，不应按正常运行时长收费；进入运行后模型不兼容、下载缓慢或自定义配置失败，应按实际占用、故障原因及交付承诺区分。第三方存储、网络和其他项目若另收费，须有事先明示和相应授权，不因页面只显示 GPU 就推定均免费或允许无限附加费用。价格变化不得任意追溯已经确认和发生的时段。
租用磁盘和运行内存不是 Laintas 长期云盘，模型、日志、输出、密钥和缓存可能随实例销毁丢失；重要结果应在停止前导出。浏览器关闭、SSH 断开或推理闲置通常不等于停止租机。余额不足、安全停机或机房故障可使任务中断，应查看实例状态与最终结算。结束时平台可清理和销毁其可控实例资源，但不保证所有机房底层介质、快照或外部备份即时物理擦除；含严格保密资料应先确认设施方和删除安排，不凭“销毁”二字推定零残留。
接入设备须由您合法控制并符合组织网络规则，保护设备令牌、管理端口与运行账户。托管转发、控制面和本机推理是不同环节；自有机器不扣模型 token 费，不表示网络流量、供电、第三方模型许可或其他服务没有成本。下载权重和部署推理需遵守相应模型的商业用途、地域和再分发限制，平台能加载不代表已取得所有许可。设备关闭或流量耗尽时应按状态失败或等待，不能据此静默改为向用户收费的其他模型。
Separate rules cover own-machine model relay and time-based GPU rental, with different charges and data responsibilities.
2026-10-25.3 · 2026-10-25
This annex applies to the relevant Laintas Compute functions together with the Platform Terms, Privacy Policy, Data Use Notice and Billing and Refund Policy. It supplements or overrides general terms only on its stated product matters. Mandatory law and valid individually negotiated agreements take priority. The version expressly provided and validly accepted at purchase or use applies; a web update does not automatically change existing orders.
Models execute on your device or chosen runtime while the gateway relays requests and answers and records bytes, timing, tokens and status metadata. This path has no Laintas per-token inference charge; membership or purchased traffic applies. Exhaustion, offline devices or unavailable allowance checks refuse or pause calls without a silent paid-model fallback. Offline status does not revoke credentials; remove the device and local credentials to end access. Power, hardware, connectivity and chosen runtime costs follow their respective agreements.
Rentals use account balance, not membership inference allowance. The GPU, hourly rate and fixed-duration or open-ended mode are shown at ordering. Current metering starts at started_at when the service successfully connects and marks the machine running, not merely on order submission. Subsequent model download and loading occupy billable time. Settlement is generally every ten minutes against actual elapsed time with a final cent-rounded reconciliation, not an automatic full ten-minute minimum per started interval. Any separately displayed fixed fee also needs confirmation. Destruction retries do not extend the user billing cutoff beyond the stop request.
Allocation failure or startup timeout before a billable running state is not normal runtime. After readiness, incompatible models or failed download/loading may still occupy rental time; stop and request cause-based review rather than assuming all rental time becomes free. Laintas-attributable non-delivery, duplicate time or misbilling requires correction and lawful remedy. Upstream faults are assessed by timing, records, order and duties, not an automatic third-party refund exclusion.
Expiry, stopping, insufficient funds, safety action or upstream interruption can stop and destroy a machine. Open-ended use is not unlimited credit; insufficient balance for the next settlement interval can trigger reclamation. Machines and disks are temporary and data is not retained after reclamation. Back up during the rental and do not use it as sole storage. Export before ending, without excluding liability for the platform’s own actionable fault. Protect SSH private keys.
Rental infrastructure may use Lium, Vast.ai or the providers actually enabled for the order. Location follows selected-machine and available order information; underlying infrastructure may be controlled by the supplier or host. You need authority over devices and model use and must comply with model, runtime, download-source and infrastructure licences and use restrictions. Downloadability does not imply resale rights. Installers may install engines, download models or enable startup as chosen; verify sources and device permissions. Private-model access tokens are limited to authorised purposes.
Current rental time starts after SSH connects successfully and the instance enters running state. Model downloads, loading and preparation then occupy billable time; charging does not wait for the first successful inference. Settlement ordinarily posts elapsed time every ten minutes and reconciles actual remaining occupancy at end, rounding to cents under ledger precision, without a ten-minute minimum per start. A successfully recorded stop request uses stop_requested_at or the relevant end time as the user cutoff. Destruction retries must not extend charges beyond that cutoff. Request a start, cutoff and debit reconciliation by rental ID if states differ.
Check GPU, memory, disk, region, availability and hourly rate at confirmation. A listing estimate is not a promise of fixed availability indefinitely. Allocation failing before normal billing state should not incur normal runtime charges. Incompatibility, slow downloads and configuration failure after running depend on occupied time, cause and delivery promises. Separately billed storage, network or other items require prior disclosure and authority; a GPU listing neither includes everything for free nor authorises unlimited extras. Rate changes must not arbitrarily reprice confirmed elapsed periods.
Rental disks and memory are not long-term Laintas Storage. Models, logs, output, keys and caches may be lost on destruction; export important results before stopping. Closing a browser, losing SSH or idle inference ordinarily does not stop rental. Insufficient balance, security shutdown and infrastructure failure can interrupt work; check instance state and final settlement. Laintas can clean and destroy controlled instance resources but does not promise instant physical erasure of every host medium, snapshot or external backup. Confirm providers and deletion arrangements before sensitive use; destruction is not a universal zero-remanence guarantee.
Connect only devices you lawfully control and within organisational network rules. Protect device tokens, management ports and runtime accounts. Relaying, control plane and local inference are distinct; no model-token charge for own-machine calls does not remove traffic, power, model-licence or other costs. Downloads and inference must respect model commercial-use, regional and redistribution conditions; technical compatibility is not a licence grant. Offline devices or exhausted traffic should fail or wait according to state, without silently switching to a paid model.