8个人，1亿美元ARR，现在还要自己训练大模型。
这不是段子，这是Base44的真实故事。一年前Wix以8000万美元收购这家成立仅半年的公司时，团队只有8个人。如今Base44正式推出自研大模型Base1，基于“平台上数千万真实用户交互数据”训练，目标是在延迟、成本和效率上全面超越调用外部前沿模型的方案。
创始人Maor Shlomo说的直白：“将模型训练纳入我们整个技术栈的一部分，让我们在延迟、成本和效率方面拥有更多优化空间。”翻译一下：调用别人的API，命脉永远在别人手里。
为什么vibe coding平台要自研模型
vibe coding——用自然语言描述需求、AI自动生成应用——是2026年最火的AI应用赛道之一。Lovable ARR达5亿美元，Base44 ARR突破1亿美元，还有大量小团队涌入。这个赛道的核心矛盾在于：产品体验建立在外部模型之上，而外部模型同时也在做自己的编程工具（Claude Code、OpenAI Codex、GitHub Copilot）。
这是一个经典的“渠道商vs品牌商”困境。你帮OpenAI和Anthropic找到了开发者使用场景，但它们随时可以通过自有产品（Claude Code、Codex）直达用户。Base44选择自研模型，本质上是在构建垂直整合的护城河：从模型到应用到用户数据，全链路掌控。
Shlomo的判断是：“模型在进步，但它们将保持通用性。”通用模型什么都能做，但在特定场景下不一定是最优解。vibe coding有自己独特的任务分布——大量前端组件生成、UI逻辑编排、数据库schema设计、部署配置——这些任务的模式高度重复，一个在这些数据上专门训练的小模型，未必比通用大模型差，但成本可以低很多。
自研模型是解药还是陷阱
但自研模型不是万能药。
Headline VC合伙人Jonathan Userovici（投资了Mistral AI）提醒不要低估前沿模型的迭代速度。他举了法律科技公司Harvey的例子——Harvey也曾计划自研模型，最终放弃。原因很简单：当你花6个月训练出一个模型，GPT或Claude的下一代可能已经把你的优势追平了。前沿模型的进步速度，可能比垂直模型的迭代速度更快。
成本也是一个现实考量。Base44坦承自研模型的成本下降“不是立竿见影”的，需要时间才能形成更强的利润结构。Wix近期宣布裁员20%，虽然Base44在被收购后持续扩张，但母公司的成本压力是真实存在的。
但Base44的赌注是：当vibe coding平台达到一定的用户规模和交互数据量，自研模型的ROI就会转正。数千万真实用户交互数据是大模型公司训练通用模型时不具备的——这些数据不是通用代码，而是真实的“人类意图→可运行应用”映射，这是vibe coding平台最核心的资产。
"企业客户并不总能在所有场景中使用最新模型获得投资回报，整个基础设施正在转向模型编排和优化。"
—— Headline VC合伙人 Jonathan Userovici
AI编程工具的战争，正在从“谁的补全更准、谁的Agent更可靠”蔓延到“谁的成本结构更可持续”。Base44选择了垂直整合，Lovable选择了继续依赖外部模型（目前ARR 5亿美元，比Base44大5倍），Cursor和Copilot则背靠大模型公司。三条路径，谁能笑到最后？
答案可能不是赢家通吃，而是分层共存——通用模型做底层基础设施，垂直模型做场景优化，应用层做用户体验。但有一点是确定的：在AI编程这个赛道，只做“套壳”的公司会越来越危险。
明天见。
8 people, $100 million ARR, and now training their own large language model.
This isn't a meme; it's Base44's real story. When Wix acquired the six-month-old company for $80 million a year ago, the team was just 8 people. Today Base44 officially launched its self-built LLM, Base1, trained on "tens of millions of real user interaction data points on the platform," aiming to surpass external frontier model solutions across latency, cost, and efficiency.
Founder Maor Shlomo put it bluntly: "Incorporating model training as part of our entire tech stack gives us more room to optimize latency, cost, and efficiency." Translation: when you call someone else's API, your lifeline is forever in someone else's hands.
Why Vibe Coding Platforms Are Building Their Own Models
Vibe coding -- describing requirements in natural language and having AI auto-generate applications -- is one of the hottest AI application tracks in 2026. Lovable is at $500 million ARR, Base44 has broken $100 million, and countless small teams are pouring in. The core tension in this赛道 is: product experience is built on external models, but those external model providers are simultaneously building their own coding tools (Claude Code, OpenAI Codex, GitHub Copilot).
This is the classic "distributor vs. brand" dilemma. You help OpenAI and Anthropic find developer use cases, but they can reach users directly anytime through their own products (Claude Code, Codex). Base44 choosing to build its own model is essentially building a vertically integrated moat: full-stack control from model to application to user data.
Shlomo's judgment: "Models are improving, but they'll remain general-purpose." General-purpose models can do anything, but aren't necessarily optimal for specific scenarios. Vibe coding has its own unique task distribution -- massive volumes of frontend component generation, UI logic orchestration, database schema design, deployment configuration -- these tasks have highly repetitive patterns, and a smaller model trained specifically on this data might not outperform a general-purpose LLM but can be dramatically cheaper.
Self-Built Models: Cure-All or Trap?
But building your own model isn't a silver bullet.
Headline VC partner Jonathan Userovici (who invested in Mistral AI) warns against underestimating frontier models' iteration speed. He cited legal tech company Harvey as an example -- Harvey also planned to build its own model but ultimately abandoned the effort. The reason is simple: by the time you spend six months training a model, the next generation of GPT or Claude may have already closed the gap. Frontier models may advance faster than vertical models can iterate.
Cost is also a real consideration. Base44 acknowledges that cost reduction from self-built models "isn't immediate" and takes time to build a stronger profit structure. Wix recently announced 20% layoffs; while Base44 has continued expanding post-acquisition, the parent company's cost pressures are real.
But Base44's bet is: when a vibe coding platform reaches a certain user scale and interaction data volume, the ROI of self-built models turns positive. Tens of millions of real user interaction data points are something LLM companies training general models don't have -- this data isn't generic code; it's real "human intent -> runnable application" mappings, the core asset of vibe coding platforms.
"Enterprise clients don't always get ROI from using the latest model in every scenario; the entire infrastructure is shifting toward model orchestration and optimization."
-- Jonathan Userovici, Partner at Headline VC
The AI coding tool wars are spreading from "whose completions are more accurate, whose Agent is more reliable" to "whose cost structure is more sustainable." Base44 chose vertical integration; Lovable chose to continue relying on external models (currently at $500M ARR, 5x Base44's size); Cursor and Copilot are backed by LLM companies. Three paths -- who laughs last?
The answer probably isn't winner-takes-all but stratified coexistence -- general-purpose models as underlying infrastructure, vertical models for scenario optimization, application layer for user experience. But one thing is certain: in the AI coding赛道, companies that only do "wrapper" products will be increasingly endangered.
See you tomorrow.
企业客户并不总能在所有场景中使用最新模型获得投资回报，整个基础设施正在转向模型编排和优化。
—— Headline VC合伙人 Jonathan Userovici
Enterprise clients don't always get ROI from using the latest model in every scenario; the entire infrastructure is shifting toward model orchestration and optimization.
-- Jonathan Userovici, Partner at Headline VC
Base44 · Vibe Coding · Self-Built Models · Cursor Competition · AI Coding Tool Vertical Integration
