热门开源AI开发者工具Ollama完成6500万美元B轮融资，月活用户近900万

热门开源AI工具Ollama已完成由Theory Ventures领投的6500万美元B轮融资，创始人兼CEO Jeff Morgan对TechCrunch表示。

此次融资紧随此前由Benchmark合伙人Peter Fenton领投的1500万美元A轮融资。至此，该公司累计融资额达8800万美元。

Ollama于2023年推出，帮助开发者在PC上运行开放权重AI模型，几分钟即可上手。它获得了无数教程网站、视频、博客和社交媒体帖子的好评。它在GitHub上已收获176,000个星标和近17,000次复刻。

开发者还可以使用Ollama查找模型，并通过从免费到每月100美元不等的多种订阅套餐，访问由其托管在新云平台上的更大、更复杂的模型。它还按GPU时间而非令牌限制来追踪用量。

如果说帮助开发者更轻松地在PC上构建应用的使命听起来有些耳熟，那确实如此。Morgan与联合创始人Michael Chiang曾参与构建Docker Desktop。在Docker收购他们之前的初创公司Kitematic后，两人加入了Docker。Docker制作的容器使得云端应用能够轻松跨云或在桌面与云之间迁移，屏蔽了所有繁琐的硬件配置问题。

因此，Ollama在AI领域本质上扮演了Docker和Docker Desktop在云领域所扮演的角色。

“开源模型在2023年开始出现，但当时它们非常难用，”Morgan表示。那时这些模型主要面向研究人员，而非程序员。“因此，要快速上手和运行它们非常困难。”推出三年后，Ollama目前“每月有超过890万开发者使用，覆盖了85%的财富500强企业，且增长势头惊人，”他说。而这一切仅靠14名员工。

正是这段职业经历吸引了Benchmark的Peter Fenton领导了其早期融资并加入董事会。

“Jeff和Michael在Docker所构建的产品每天被超过1000万开发者使用。能够打造出在开发者群体中无处不在的产品的创造力极为罕见，”Fenton对TechCrunch表示。

Morgan和Fenton拒绝对该初创公司的营收和新估值置评。不过，Morgan表示，Ollama商业模式的验证时刻发生在今年1月左右，那时OpenClaw变得火爆。正是那时，更大的开源模型“突然能够执行诸如编程等智能体任务。显然，我们看到了像OpenClaw这类助手的爆发，以及开源模型能够完成实际工作的理念。”

自那以后，业界热议着一个观点：付费用户（尤其是财力雄厚的企业和快速增长的人工智能应用层初创公司）将越来越多地转向更经济实惠的开源模型，而将像Anthropic那样的闭源模型留作按需使用。

“我仍认为，这是多数争论的误区。这不是非此即彼的问题，”Fenton在谈及开源与闭源AI模型时表示。他主张，两类模型都会有大量商机。不过，每条推理费用高昂——即使用模型的成本——的公司都有一项“攸关存亡的项目”推动他们转向“开放权重模型”。

有大量证据表明，这类初创公司和企业已经在日常使用中转向开源模型。这显然对Ollama的云业务是个好兆头。

但更有趣的是，Ollama是AI催生大量新开源项目并转化为VC追捧公司的又一例证。有诸如vLLM的创建者Inferact和SGLang的创建者RadixArk这样的开源推理提供商。还有OpenClaw及其替代品如NanoClaw。甚至还有像Arcee这样从零开始构建自有开源模型的小型初创公司。

诚然，并非所有Ollama粉丝都对该公司谋求商业化感到高兴。大约一年前，一系列博客和社交媒体帖子抱怨其云业务分散了对备受喜爱的免费项目的关注，并引用Ollama作为开发者工具所谓“平台腐化”现象的例证。

但Morgan将其云服务视为其开源使命的演进，即帮助程序员发现并轻松使用模型。那些先进的大型开源模型通常“太大而无法在自己的电脑上运行。所以我们说，‘嘿，我们来帮忙寻找算力吧，’”他解释道。

董事会成员Fenton补充说：“桌面版免费核心产品没有任何改变。关于这里是发现和运行本地模型之地的承诺，也没有丝毫变化。”

The popular open source AI tool Ollama has raised a $65 million Series B, led by Theory Ventures, founder and CEO Jeff Morgan tells TechCrunch.

This round follows a previous $15 million Series A led by Benchmark’s Peter Fenton. All told, the company has now raised $88 million.

Ollama, which launched in 2023, helps devs run open-weight AI models on their PCs, getting them up and running in minutes. It has been praised by developers across countless training sites, videos, blogs and social media posts. It has amassed 176,000 stars and nearly 17,000 forks on GitHub.

Developers can also use Ollama to find models and access larger, more complex ones that it hosts on its neocloud via several subscription tiers, from free to $100/month. It also tracks usage based on GPU time, not token limits.

If the mission to help developers more easily build on their PCs sounds vaguely familiar, it should. Morgan and his co-founder Michael Chiang previously helped build Docker Desktop. They landed at Docker after it bought their previous startup, Kitematic. Docker makes containers that help cloud apps easy to move from cloud to cloud, or from desktop to cloud, abstracting away all the pesky hardware configuration issues.

So Ollama essentially did for AI what Docker and Docker Desktop did for cloud.

“Open models started coming out in 2023 but they were really hard to use,” Morgan said. They had been geared toward researchers at the time, not programmers. “As a result, it was really hard to get them up and running.” Three years after launching, Ollama is now “used by over 8.9 million developers every month, sitting in 85% of the Fortune 500 and growing like crazy,” he said. All with only 14 employees.

That career experience is what drew Benchmark’s Peter Fenton to lead its earlier round and join the board.

“What Jeff and Michael built with Docker is being used by 10 million-plus developers every day. The creative powers to create a product that goes to ubiquity for developers is extremely rare,” Fenton told TechCrunch.

Morgan and Fenton declined to discuss the startup’s revenues and new valuation. However, Morgan says that the proving point for Ollama as a business happened around January, when OpenClaw became hot. That’s when larger open models “suddenly became able to do these agentic tasks, like coding. Obviously, we saw the explosion of the assistants like OpenClaw, and this idea that open models can get real work done.”

Since then, the industry has been abuzz with the idea that paying users (particularly deep-pocketed enterprises and fast-growing AI application-layer startups) will increasingly turn to more affordable open models, reserving their use of closed models like Anthropic for more of an as-needed basis.

“I still think that this is the part that most of the debate gets wrong. It’s not an either/or,” Fenton says of open versus closed AI models. There will be plenty of business for both, he contends. However, every company with high inference expenses — the costs of using the models — has a “vital existential project” pushing them to move “to open-weight models,” he says.

There’s plenty of evidence that such startups and enterprises are already turning to open models for their daily needs. That, obviously, bodes well for Ollama’s cloud business.

But even more interesting, Ollama is another example of how AI is birthing a large new crop of open source projects that are turning into companies pursued by VCs. There are open source inference providers like Inferact, maker of vLLM, and RadixArk, maker of SGLang. There is OpenClaw and its alternatives like NanoClaw. There are even tiny startups building their own open models from scratch, like Arcee.

To be sure, not every Ollama fan has been happy that the company has been pursuing making a living. About a year ago, a bunch of blog and social media posts complained that its cloud business was drawing attention away from its beloved free project and cited Ollama as an example of the so-called “Enshittification” of dev tools, as the trend is called.

But Morgan sees its cloud service as an evolution of its open source mission to help programmers find and easily use models. Those state-of-the-art, large, open models are often “too big to run on your own computer. So we said, ‘Hey, let’s help find the compute for that,’” he explained.

Board member Fenton adds, “Nothing has changed for the core product that’s free on the desktop. There’s zero change to the premise that this is the place you can discover and run local models.”