AI 模型 API 平台AI model API platform

从算子做起的
Token 平台
The token platform
built from the kernel up

TokenFly 是通用 Token 平台,一个接口调用文本、图像、语音和视频模型。技术底座由清华技术团队打造,核心技术是基于 RSI 的自演进算子优化。TokenFly is a general token platform: one API for text, image, audio and video models. It is built by a Tsinghua engineering team, on RSI-based self-evolving operator optimization.

162
可调用模型models available
22
模型厂商model makers
—
对比厂商官方价against list price
热门模型Popular models文本:输出价 / 百万 tokensText: output / 1M tokens
deepseek-v4.1-flash77% off$0.14$0.60
deepseek-v4.1-flash-101777% off$0.14$0.60
glm-5.368% off$1.39$4.40
glm-5.3-flash56% off$0.22$0.50
minimax-h3768P · 42% off$0.05/sec$0.08
TokenFly 价TokenFly price厂商官方价List priceOpenAI 格式接口的实时价格,其他接口见模型价格页Live OpenAI-format prices; other interfaces are on the model page

高质量High quality

接口兼容主流调用方式,工具调用、流式和 JSON 模式照常使用。自研优化同时检查速度和质量,不达标的方案不会上线。Compatible with the mainstream APIs, so tools, streaming and JSON mode work as before. Our own optimizations are checked for speed and quality, and those that fail never ship.

低价格Low prices

输入和输出单价都不高于厂商官方价。按实际用量计费,每笔费用都能在控制台查看。用量大可以签约获得专属价格和额度。Input and output prices never exceed the maker's list price. You pay for what you use, and every charge is in the dashboard. High-volume customers can sign for dedicated prices and capacity.

清华技术团队Tsinghua engineering

技术底座由清华技术团队打造,研究方向覆盖大模型、强化学习和推理系统,并把研究成果落到线上服务。Built by a Tsinghua team working across large models, reinforcement learning and inference systems, with the research running in production.

核心技术:
基于 RSI 的自演进算子优化
Core technology:
RSI-based self-evolving operators

RSI 即递归自我改进。智能体持续提出并实现算子优化,自动门控检查质量和速度。通过的代码与经验成为下一轮的起点。目前已在 330 亿参数视频模型上线验证。RSI is recursive self-improvement. Agents keep proposing and implementing operator optimizations, an automatic gate checks quality and speed, and what passes becomes the next round's starting point. It is validated in production on a 33B-parameter video model.

  1. 提出并实现Propose and build

    智能体定位瓶颈,改写算子和执行路径。Agents find bottlenecks and rewrite operators and execution paths.

  2. 自动门控Automatic gate

    与官方实现做数值对齐,再测速度。Numerics are checked against the reference, then speed is measured.

  3. 存档与继承Archive and inherit

    只保留通过检验的代码和经验报告。Only code and reports that passed are kept.

  4. 人工终审Human sign-off

    发布前评审,不参与优化反馈。Reviewed before release, outside the optimization loop.

算子优化的实测加速Measured operator speedups

MiniMax-H3,单次前向或单项对比MiniMax-H3, single forward pass or single comparison

Sol 稀疏注意力Sol sparse attention2.62× NVFP4 W4A41.37× 融合 RMSNormFused RMSNorm1.19× 逐点算子融合Pointwise fusion1.14×

各项为单独测量,刻度 0–3×。Measured separately; scale 0–3×.

更快但不达标的方案,我们拒绝了Faster, but rejected

门控记录节选From the gate log

候选方案Candidate未通过的原因Why it failed
融合算子,快 7 倍Fused operator, 7× faster视频与音频的数值偏差超标Video and audio numerics out of tolerance
8 步加速 LoRA8-step acceleration LoRA画质指标未达标Image quality below target
编译版 VAECompiled VAE输出 148 帧,应为 124 帧Output 148 frames instead of 124
案例:330 亿参数视频模型,约 7 秒交付Case: a 33B video model, delivered in about 7 seconds 1344×768 · 124 帧 · 3 步 · 8 卡服务器 · 线上中位数 n=131344×768 · 124 frames · 3 steps · 8-GPU server · production median, n=13
2.81 s去噪完成Denoised
5.17 s成片时长Clip length
6.96 s完整请求,含封装Full request, packaged
02468 s
去噪Denoising生成完成 5.23 sGenerated at 5.23 s封装与发布Packaging
Claude Code · Codex

一行命令,切换到 TokenFlyOne command to switch to TokenFly

正在加载…Loading…
查看脚本View script 获取 API KeyGet an API key

粘贴到终端回车,输入 API Key(不显示)后运行Paste into a terminal, enter your API key (hidden), then run claude。适用于 macOS / Linux,Windows 请用 WSL。. macOS / Linux; on Windows use WSL.

告诉我们你的业务规模Tell us your scale

带上你的模型和预估用量,我们会先判断是否适合,再安排样本测试。Bring your models and expected volume. We will check the fit first, then set up a sample test.

微信联系Contact on WeChat 查看二维码Show QR code
微信联系二维码 / WeChat contact QR code

使用微信扫一扫联系,或保存图片后在微信中识别。Scan with WeChat, or save the image and open it in WeChat.

查看原图Open full-size image
飞书联系Contact on Feishu 查看二维码Show QR code
飞书联系二维码 / Feishu contact QR code

使用飞书扫一扫,添加联系人;也可保存二维码后在飞书中识别。Scan with Feishu to add us, or save the image and scan it in Feishu.

查看原图Open full-size image

只想先试试?Just want to try it? 注册并获取 API KeySign up for an API key