fim.ai

mx

mx1.feishu.cn

mx2.feishu.cn

mx3.feishu.cn

ns

curitiba.ns.porkbun.com

salvador.ns.porkbun.com

maceio.ns.porkbun.com

fortaleza.ns.porkbun.com

caa

txt

verification-code-site-App_feishu=tS3lzS2S5dFICtsvGGEf

google-site-verification=Tt9ky6WI0ivXypW4EX_x-eA51MuhY572s-qVGeZvBxc

v=spf1 +include:_netblocks.m.feishu.cn -all

google-site-verification=_bd5_BC3ZxzCWzGIGmxq4bzkLlLRa0HEefBD91zIC9Y

dmarc

User-Agent: *
Allow: /
Disallow: /api/

User-Agent: GPTBot
User-Agent: OAI-SearchBot
User-Agent: ChatGPT-User
User-Agent: ClaudeBot
User-Agent: Claude-SearchBot
User-Agent: Claude-User
User-Agent: PerplexityBot
User-Agent: Perplexity-User
User-Agent: Google-Extended
User-Agent: Applebot-Extended
User-Agent: Baiduspider
User-Agent: Bytespider
User-Agent: Sogou web spider
User-Agent: 360Spider
User-Agent: YisouSpider
User-Agent: PetalBot
Allow: /
Disallow: /api/

Sitemap: https://fim.ai/sitemap.xml

# FIM (飞沐 / FIM Labs)

> FIM is an AI product company with two product lines. Zico is a self-hosted contract lifecycle management platform, running in production at groups such as Newborn Town. One is a model-agnostic enterprise agent platform, source-available and self-hostable. Around our research work we also build a set of research tools: Cito, Tuto, Aha., and Hido.

FIM operates through two companies: FIM Labs Pte. Ltd. in Singapore and Beijing FIM Network Technology Co., Ltd. in Beijing (founded 2013). The two are independent of each other and share the FIM brand.

中文：FIM 是一家 AI 产品公司，当前有两条产品主线：Zico 是私有化部署的合同全生命周期管理平台，已在赤子城等集团的生产环境运行；One 是模型无关的企业级 Agent 平台，源码可见、可自托管。围绕研究工作，我们还做了 Cito、Tuto、Aha.、Hido 等研究工具。

FIM 在新加坡与北京各有一家公司：FIM Labs Pte. Ltd. 与北京飞沐网络科技有限公司（2013 年成立）。两家公司相互独立，共用 FIM 品牌。

## Flagship products (https://fim.ai/en/products · 中文：https://fim.ai/zh/products)
- [One](https://one.fim.ai): Model-agnostic enterprise agent platform. Dynamic DAG planning, a RAG retrieval pipeline, and a visual workflow editor. Works with any OpenAI-compatible model and deploys on-prem. (Source-available · GitHub)
- [Zico](https://zico.fim.ai/): Self-hosted contract lifecycle management. Deep integration with your collaboration platform (Feishu / WeCom / DingTalk / major OA, pick one), AI contract-risk review, and the full drafting-to-signing-to-archive flow, with private deployment. (In production at Newborn Town and other groups)

## Research tools
- [Cito](https://cito.fim.ai): Semantic paper search. Search papers by meaning, not keyword overlap: SPECTER2 fused with BM25 retrieval, covering ~148M papers. (Open API · Docs)
- [Tuto](https://tuto.fim.ai): Citation-integrity auditing. Checks two layers of a citation: whether the cited work exists, and whether it supports the claim. First run: all 209,985 references in ACL 2026. (Open dataset · Report)
- [Aha.](https://aha.fim.ai): Plain-language arXiv paper explainer. Paste an arXiv link and read the paper side by side with a plain-language explanation.
- [Hido](https://hido.fim.ai): Anonymous code hosting for double-blind review. Turn a GitHub repo into a read-only mirror that does not leak author identity. Reviewers browse, download, and clone with no login. One anonymization pipeline, prerendered to R2 for static serving. (Console live)

## Open learning resources
- [Build GPT-2 from Scratch](https://learn-llm.fim.ai/): Hands-on course: write your own GPT-2. A Mandarin course that rebuilds GPT-2 end to end, from tokenization to training, one step at a time.
- [Linear Algebra → Attention](https://l2a.fim.ai/): Visual path from linear algebra to attention. Interactive visualizations that walk from linear algebra all the way to the Transformer attention mechanism.
- [Algorithm Visualization Lab](https://algo.fim.ai/): Drag to see the geometry behind algorithms. Interactive algorithm visualizations: drag a handle, watch a sort or a graph traversal rearrange, and read off the geometric idea underneath. A Chinese living notebook that grows one algorithm at a time.

## Services (https://fim.ai/en/solutions · 中文：https://fim.ai/zh/solutions)
FIM delivers enterprise AI built on One and Zico: private deployment, agents and knowledge-base Q&A, system integration, and corporate AI training. The Deployments page lists 8 delivered projects across enterprise, healthcare, and academia.

中文：FIM 承接企业 AI 落地交付：私有化部署、智能体与知识库问答、与现有系统集成，以及企业内训与岗位上岗，底座是自研的 One 与 Zico。落地案例页列出 11 个已交付项目，分布在企业、医疗、高校与政务。

- [Corporate AI Training](https://fim.ai/en/solutions#training): Day-rate workshops for teams. No purchase commitment. Leadership sessions on limits; hands-on workflows for business and engineering.
- [Custom Solutions](https://fim.ai/en/solutions#custom): Private deployment and custom build. Built on One and Zico against your data, systems, and compliance requirements.

### China-only services (https://fim.ai/zh/solutions)
- [Rolo 岗位上岗包](https://rolo.fim.ai/): 五天完成一个岗位的 AI 上岗。固定范围的「人 + AI + 表格」工作流。资产可带走，不绑定 One。
- [Cavo 安全服务](https://cavo.fim.ai/): 面向网站与服务器的周期体检。全程只读采集；结论经公开漏洞库核实后写入报告。

## Deployments (https://fim.ai/en/deployments · 中文：https://fim.ai/zh/deployments)
- Newborn Town Inc.: The whole contract flow in one system (Zico). Drafting, AI risk review, signing, and archiving never leave the system, natively integrated with Feishu.
- A national mobility services group: Contract system bound to the internal OA (Zico). Drafting, risk review, signing, and archiving connected, with approvals flowing through Feishu.
- Tsinghua University: Student writing and teacher grading on one platform (One). A platform built on One provides real-time text analysis, grammar and logic feedback, AI-assisted drafting, and teacher-assisted grading.
- Peking Union Medical College Hospital: Specialty experience captured as a structured graph. A structured gynecology knowledge graph carries specialty clinical knowledge for decision-support and assisted diagnosis.
- Peking Union Medical College Hospital: Surgical approach recommended per patient. A model-driven recommender supports pelvic-organ-prolapse (POP) decisions, factoring in patient-specific conditions.
- Beijing Ditan Hospital / WHO Collaborating Centre: Drug lookup and patient follow-up in one platform. The specialty follow-up platform covers HIV drug-interaction lookup and a patient education portal (Jingshun No. 8).
- China Cancer Foundation: Grassroots radiotherapy teams train and certify online. The platform serves radiation oncologists, physicists, and technologists with structured courses, an examination center, and official certificates.
- Peking University: Quantitative paleontology in the classroom. The platform handles quantitative analysis and visualization of paleontological data, used directly in coursework by faculty and students.

### China-only deployments (https://fim.ai/zh/deployments)
- 政府法律 / 法律援助：规范性文件先过一遍智能体审查 (One)。One 智能体以专用模型加 RAG 完成合规与逻辑一致性审查，并提供法律援助问答与案情初判。
- 党史研究室：从史料检索到初稿走通一条流程 (One)。AI 辅助编纂系统承担资料检索、结构化组织与初稿生成，覆盖党史与地方志的编纂流程。
- 海关系统：贸易趋势与进口合规并到一个平台。平台以大数据监测技术性贸易趋势，AI 物种识别与命名库支撑进口合规审核。

### Case studies
- [Newborn Town Inc.: The whole contract flow in one system](https://fim.ai/en/deployments/newborn-town-zico) · 中文：https://fim.ai/zh/deployments/newborn-town-zico
- [A national mobility services group: Contract system bound to the internal OA](https://fim.ai/en/deployments/services-group-zico) · 中文：https://fim.ai/zh/deployments/services-group-zico

## Research (https://fim.ai/en/research · 中文：https://fim.ai/zh/research)
- AI as Equalizer or Amplifier? Task Complexity as the Moderating Factor for Human Expertise in Hybrid Intelligence Systems. Tao An. HHAI 2026 · IOS Press, 2026. [arXiv](https://arxiv.org/abs/2512.10961) · [IOS Press](https://ebooks.iospress.nl/doi/10.3233/FAIA260506) · [Aha.](https://aha.fim.ai/paper/2512.10961v2) · [Interactive note](https://tao-hpu.github.io/articles/equalizer-amplifier) · walkthrough: https://fim.ai/en/articles/ai-equalizer-or-amplifier · 中文导读：https://fim.ai/zh/articles/ai-equalizer-or-amplifier
- When Should the Agent Speak? A Survey of Intervention Timing for Always-On AI Assistants. Tao An. TMLR 2026 · Survey, 2026. [OpenReview](https://openreview.net/forum?id=b0yKEdAXEr) · [Preprint (Zenodo)](https://doi.org/10.5281/zenodo.21438396) · [Interactive note](https://tao-hpu.github.io/articles/intervention-timing) · [Living map](https://github.com/tao-hpu/awesome-proactive-agents) · walkthrough: https://fim.ai/en/articles/when-should-the-agent-speak · 中文导读：https://fim.ai/zh/articles/when-should-the-agent-speak
- Consensus Density Predicts Output Dispersion in Aligned LLMs. Tao An, Shuai Feng. TMLR 2026 · Accepted, 2026. [OpenReview](https://openreview.net/forum?id=6ukieTMBcG) · [GitHub](https://github.com/tao-hpu/consensus-dispersion) · [Interactive note](https://tao-hpu.github.io/articles/consensus-dispersion) · walkthrough: https://fim.ai/en/articles/preference-centroid-dispersion · 中文导读：https://fim.ai/zh/articles/preference-centroid-dispersion
- Recombination or Discovery? A Retrieval-Grounded Novelty Audit of Machine-Generated Research Papers. Tao An. NeurIPS 2026 Workshop · AI4MetaScience · Poster (non-archival), 2026. [Preprint (Zenodo)](https://doi.org/10.5281/zenodo.21696223) · [GitHub](https://github.com/tao-hpu/machine-science-audit) · [Interactive note](https://tao-hpu.github.io/articles/novelty-audit) · walkthrough: https://fim.ai/en/articles/recombination-or-discovery · 中文导读：https://fim.ai/zh/articles/recombination-or-discovery
- Cognitive Workspace: Active Memory Management for LLMs. Tao An. Preprint · arXiv, 2025. [arXiv](https://arxiv.org/abs/2508.13171) · [Aha.](https://aha.fim.ai/paper/2508.13171v1) · [Interactive note](https://tao-hpu.github.io/articles/active-memory-revisited) · walkthrough: https://fim.ai/en/articles/cognitive-workspace-beyond-rag · 中文导读：https://fim.ai/zh/articles/cognitive-workspace-beyond-rag
- Fidelity Before Structure: Verbatim Chunks Beat Lossy Artifact Extraction in Long-Conversation LLM Memory. Tao An. Preprint · arXiv, 2026. [arXiv](https://arxiv.org/abs/2601.00821) · [Aha.](https://aha.fim.ai/paper/2601.00821) · [Dataset](https://huggingface.co/datasets/tao-hpu/cog-canvas-benchmark) · [Interactive note](https://tao-hpu.github.io/articles/verbatim-memory) · walkthrough: https://fim.ai/en/articles/verbatim-chunks-llm-memory · 中文导读：https://fim.ai/zh/articles/verbatim-chunks-llm-memory
- Registers, Not Plans: What Lives in a Language Model's Workspace That Isn't on Its Tongue. Tao An. Working paper, 2026. [GitHub](https://github.com/tao-hpu/jspace-replication) · [Interactive note](https://tao-hpu.github.io/articles/workspace-registers) · walkthrough: https://fim.ai/en/articles/registers-not-plans · 中文导读：https://fim.ai/zh/articles/registers-not-plans
- CGEP: Toward Detecting and Attributing GEO Poisoning in Chinese AI Search. Tao An. Working paper, 2026. walkthrough: https://fim.ai/en/articles/cgep-geo-poisoning-chinese-ai-search · 中文导读：https://fim.ai/zh/articles/cgep-geo-poisoning-chinese-ai-search
- A Graph-Neural-Network Method for Data-Information Recommendation. Tao An. Patent · Under examination, 2026.

## Articles (paper walk-throughs & practice notes)
- [Are Machine-Written Papers More Novel? The Gap Sits in One Facet](https://fim.ai/en/articles/recombination-or-discovery): A read of a NeurIPS 2026 AI4MetaScience workshop poster (non-archival): machine-generated papers are judged facet-novel more often than matched human papers, but the gap comes mainly from the purpose facet, and LLM re-auditors return refutation rates from 0% to 100% on the same verdicts. · 中文：[机器写的论文更「新」吗？差距主要在「目的」一个维度](https://fim.ai/zh/articles/recombination-or-discovery)
- [Equalizer or Amplifier? The Dividing Line Is Task Complexity](https://fim.ai/en/articles/ai-equalizer-or-amplifier): A read of an HHAI 2026 position paper: AI narrows the novice-expert gap on routine work and widens it on complex work. What decides who benefits is domain expertise, not prompt engineering. · 中文：[AI 是均衡器还是放大器？分水岭在任务复杂度](https://fim.ai/zh/articles/ai-equalizer-or-amplifier)
- [Beyond RAG: Giving a Language Model Working Memory It Actually Manages](https://fim.ai/en/articles/cognitive-workspace-beyond-rag): A read of the arXiv preprint on Cognitive Workspace, which replaces retrieve-and-forget with active memory management and layered cognitive buffers: 58.6% memory reuse and a 17–18% net efficiency gain. Includes a self-correction written a year later. · 中文：[超越 RAG：给大模型一个会主动管理的工作记忆](https://fim.ai/zh/articles/cognitive-workspace-beyond-rag)
- [When Should an Always-On Assistant Speak Up?](https://fim.ai/en/articles/when-should-the-agent-speak): Survey walkthrough (published in TMLR, 2026). Research on interruptibility from 1999 formalized the cost of interrupting all the way down to utility theory. Proactive agents built since 2024 dropped that term entirely. The survey reconnects the two lines under a single decision rule. · 中文：[常驻 AI 助手什么时候该开口？](https://fim.ai/zh/articles/when-should-the-agent-speak)
- [What Actually Lives in a Model's Workspace: Registers, Not Plans](https://fim.ai/en/articles/registers-not-plans): Paper notes: an independent replication of the "global workspace" claim on open-weight models. Once you filter using the model's own next-token distribution, the only thing left that survives causal manipulation is context registers. Content plans fall to the permutation baseline. · 中文：[大模型的工作空间里住着什么：寄存器，不是计划](https://fim.ai/zh/articles/registers-not-plans)
- [Ask the Same Question Ten Times: Consensus Density Predicts How Much the Answers Spread](https://fim.ai/en/articles/preference-centroid-dispersion): Paper walkthrough (accepted at TMLR 2026): the output dispersion of an aligned model is predicted by the consensus density of the task. A judge that sees only the prompt, never the outputs, predicts dispersion at rho = -0.91, and alignment turns out to amplify a gradient the base model already had. · 中文：[同一个问题问十次，答案有多散？看任务的共识密度](https://fim.ai/zh/articles/preference-centroid-dispersion)
- [Long-Conversation Memory: Raw Transcript Chunks Beat Structured Extraction](https://fim.ai/en/articles/verbatim-chunks-llm-memory): Paper notes: a controlled ablation inside a fixed pipeline finds that storing raw conversation chunks outscores LLM-extracted structured memory by 15.9 points on LoCoMo and 22.0 points on LongMemEval-S, at a lower cost per correct answer. · 中文：[长对话记忆：原文切片赢了结构化提取](https://fim.ai/zh/articles/verbatim-chunks-llm-memory)
- [Content Poisoning in Chinese AI Search: Detecting It Is Not the Same as Knowing Who Did It](https://fim.ai/en/articles/cgep-geo-poisoning-chinese-ai-search): Paper walkthrough: detection and attribution of GEO poisoning against DeepSeek, Doubao and Kimi. Content features catch the manipulation (F1 0.93) but cannot name the actor; account-interaction graphs do the opposite (0.96). The two need separate models. · 中文：[中文 AI 搜索的内容投毒：检测到了，不等于知道是谁干的](https://fim.ai/zh/articles/cgep-geo-poisoning-chinese-ai-search)
- [The Truth About AI Right Now: An Expert Interview Without the Hype](https://fim.ai/en/articles/the-truth-about-ai-right-now-an-expert-interview-without-the-hype): AI is everywhere right now — on roadmaps, in boardrooms, and across product conversations. But once you look past the headlines, the questions that matter are still very practical: What’s actually working in real deployments?  · 中文：[AI的真相：一场没有炒作的专家访谈](https://fim.ai/zh/articles/the-truth-about-ai-right-now-an-expert-interview-without-the-hype)

## Site
- Home: https://fim.ai/en · 中文：https://fim.ai/zh
- Products: https://fim.ai/en/products · 中文：https://fim.ai/zh/products
- Solutions: https://fim.ai/en/solutions · 中文：https://fim.ai/zh/solutions
- Deployments: https://fim.ai/en/deployments · 中文：https://fim.ai/zh/deployments
- Research: https://fim.ai/en/research · 中文：https://fim.ai/zh/research
- Articles: https://fim.ai/en/articles · 中文：https://fim.ai/zh/articles
- About: https://fim.ai/en/about · 中文：https://fim.ai/zh/about
- Careers: https://fim.ai/en/careers · 中文：https://fim.ai/zh/careers
- Contact: https://fim.ai/en/contact · 中文：https://fim.ai/zh/contact · hi@fim.ai
- GitHub: https://github.com/fim-ai
- LinkedIn: https://www.linkedin.com/company/fim-labs
- Full text for language models: https://fim.ai/llms-full.txt