文 | 字母AI
训练一个5000亿参数的大模型，需要多少人？
AI行业的经验是，至少几十人，大厂动辄投入数百人。
但硅谷一家叫Mirendil的初创公司，把这个数字压缩到了极致——只要一个人。
就在刚刚，Mirendil在官网发了一篇文章，标题非常耸人听闻：《无需预训练团队的大规模预训练》。
文章标题下面列着三个数字：
1名研究员、三个半月、成本不到1000万美元。
这些钱涵盖了Agent的token消耗、数据处理和全部研究实验成本。只靠这些投入，他们真的把千亿参数的模型训出来了：Mirendil-Base，一个1000亿到 5000亿参数区间的编码专用基座模型。
Mirendil在文章中称，新模型超过99%的代码不是人写的，而是Agent写的。
人类研究员只需要负责确定方向、下判断，在关键节点踩刹车。写代码、跑实验、错误排查之类的活儿，全交给了Agent。
在Mirendil的测试中，新模型的部分指标已经逼近了Kimi K3、GLM-5.3、DeepSeek-V4等主流一线模型。
“单人手搓基模”的时代，真的要到来了……吗？
01
按照官方说法，Mirendil-Base是Mirendil系列模型的首发产品，定位是编码专用基座模型，目标用户是前沿AI实验室内部的研究员和工程师。
为了验证一个人究竟能不能把模型做出来，Mirendil设置了许多苛刻条件。比如，不能用现成的预训练框架，也不能继承代码。
研究员和Agent从零起步，在100多天里搭建了一切——训练基础设施、模型架构、训练算法、数据管道，全是自己写的，只是在PyTorch、JAX这几个通用库上，参考了一些热门开源项目当样板。
当然，这里的“自己”，通常指的是Agent。
研究员主要负责决定尝试什么、信任什么，并在需要时干预。写代码、跑实验、诊断报错这些具体的动作，基本都是Agent自己跑。
数据方面，Mirendil全部使用公开数据，比如GitHub开源代码、公开网页、公开数据集之类。公司强调，训练过程中没买过一行数据，没做过蒸馏，也没用合成数据。
此外，Mirendil希望通过改进设计提升效率。“我们通过改进的架构、数据混合的高级调优，以及其他数据层面的改进，还有kernel优化，获得了显著的效率提升。”公告称。
最终成品是Mirendil-Base，参数为1000 亿～5000 亿。
一个人做出来的千亿参数模型，到底是个什么水平？
Mirendil拿了五款模型做对比，分别是Kimi K3、GLM-5.3、Nemotron-550B-Base、DeepSeek-V4-Pro-Base、Inkling，均为MoE架构的大参数模型。
在新代码仓库的交叉熵测试（模型猜测代码怎么写的，分数越低、猜得越准）中，几款模型的比拼结果如下：
Kimi K3排第一，得分0.193；Mirendil-Base排第二，得分0.222；后面依次是GLM-5.3、DeepSeek-V4、Nemotron。
换句话说，在部分编码维度上，一个人和Agent组队训练出来的预训练模型，已经能跟头部大厂的成品模型掰手腕了。
不过，必须指出的是，通常情况下，基座模型在交叉熵测试中比后训练模型表现更好。Mirendil-Base作为基座模型，分数反而比后训练过的Kimi-K3更高，说明 Kimi-K3的真实实力比表面分数显示的还要强得多，Mirendil-Base跟它的差距比看起来更大。
根据官方介绍，Mirendil-Base往上还有两档：Large在5000亿到2万亿，XL直接冲到2万亿到10万亿。
目前，Large训练了80%，XL刚开了个头，才19%，却已经快追上Large现在的分数了。
也就是说，在钞能力和Agent的加持下，一个人训练万亿模型不是梦——这是许多AI大厂都还没摸到的门槛。
02
千亿乃至万亿参数大模型的预训练，极为耗钱、费时，还必须大量堆人。
几百位研究员里面，一部分研究模型该长什么样、怎么做scaling；一部分人管数万亿token 的清洗和配比；还有一部分人守着几千张，甚至几十万张算力卡，保证训练不停机。
AI大厂可以聚拢足够多的人才，但资金和算力成本压力依然很大。
整个行业一直在尝试降本，核心思路就是一条：让AI来训练AI。
比如，Anthropic已经沿着这条路走了很远。
它此前披露，今年5月，其代码库里超过80%的合并代码是Claude写的。
谷歌也在做类似的事情：AI编程智能体AlphaEvolve，已经被用来优化数据中心调度、芯片设计和AI训练流程。
但“AI训练AI”这条路，目前只有大厂才能玩得转。
科学家、研究员、小团队等，想自己训练一个模型，以前几乎不可能。
Mirendil-Base却证明了，在一定条件下，一个人也可以“手搓”模型。
Mirendil在官网文章中称，直到不久前，预训练一个有竞争力模型的瓶颈还是人：一支需要多年才能搭建起来的专家团队。
如今，这一局面正在改变。
过去只有大型研究机构才能做到的事，现在一名研究员就可以完成。当Agent承担了大部分动手工作时，瓶颈转移到了算力、数据，以及指挥Agent的那个人的判断力上。
“人人都能训练模型”，蕴含着巨大的想象空间。
以前，大多数人和公司如果需要专属模型，往往租用一个通用模型、再在边缘做适配。
现在，一支芯片设计团队、一个机器人实验室，或是一家医院，都可以用自己的数据、为自己的问题训练一个模型。基座模型以及围绕它们构建的系统，将不再是只有少数大厂才能制造的东西，而会更加普及。
另一方面，过去需要一个大型团队才能启动的AI项目，现在可以从小团队，甚至一个人起步；而随着算力越来越便宜、越来越强大，更多人能够去尝试这些项目。“一人公司”或者“超级个体”，有望变成现实。
人们的工作方式也会被改变：随着越来越多的工作被交给Agent，人类的日常工作将越来越不像“亲手做事情”，而越来越像“带领一个小型实验室”：设定研究方向、设计值得运行的实验、判断哪些想法值得继续推进。
Mirendil描绘了更加炫目的前景：下一个模型将部分由Mirendil自己的模型来构建，而每一个在构建 AI这件事上变得更强的模型，都会让下一个模型的构建更便宜、更快速。
“我们认为，还没有达到完全自主的RSI，但我们已经比以往更接近了。”公司称。
此外，公司已经开始对Mirendil-Base进行后训练。Mirendil-Base-Large以及达到前沿水平的Mirendil-Base-XL仍在训练中。
03
Mirendil的背后，是两位Anthropic前员工：Behnam Neyshabur和Harsh Mehta。
Neyshabur是80后，伊朗出生、读大学。2011 年去了美国，在丰田工业大学芝加哥分校拿了博士学位。毕业后，他在谷歌DeepMind干了5年，2024年底加入Anthropic，领导了Discovery团队，专门研究怎么让AI像科学家一样参与研发。
Mehta是印度裔，印度理工学院古瓦哈提分校毕业的，在DeepMind的时候跟Neyshabur是同事。加入Anthropic之后，Mehta搭建了Anthropic内部自动化AI研究平台的第一个版本。
2025年12月，两人前后脚离职，成立了Mirendil，公司定位是让AI来帮助科学家造出他们自己的AI，而不只是直接用AI辅助科学。
换句话说，他们希望让科研团队能够自己训练、迭代专属AI模型，不用再仰仗大厂的通用模型。背后的技术路线，就是最近AI圈内炒得很热的RSI。
今年6月，Mirendil官宣了 2 亿美元种子轮融资，投后估值 10 亿美元。投资方包括a16z、凯鹏华盈、英伟达等。
团队目前约 20 人，都是从Anthropic、xAI、Google DeepMind、OpenAI 出来的研究员和工程师，在旧金山市中心办公。
Mirendil-Base的亮相，标志着公司跨过了第一个里程碑。
但话说回来，“单人训练千亿参数模型”听起来美好，门槛并没有被拉低多少。
1000万美元的成本，对大厂来说不算什么，刚刚拿到融资的Mirendil也还付得起。但对于绝大多数个人、小团队来说，依然是天文数字。
更别说，能指挥 Agent整出如此大活儿，不是随便拉一个工程师就能上手的。Mirendil的人才质量之高、AI项目经验之丰富，在整个硅谷都处于第一梯队。
“一人手搓大模型”的时代确实在靠近，但离普通人还很远。
Mirendil-Base 的发布，是这条漫长道路的第一块路标。至于它能不能真的把大模型的门槛打下来，还得看后续更大规模模型的实际水平，以及这套方法论能不能复制到别人身上。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Letter AI
How many people do you need to train a LLM of 500 billion parameters?
The experience of the AI industry is that there are at least a few dozen people and hundreds of people in large factories.
But a start-up company in Silicon Valley, Mirendil, condensed that number to the extreme — just one person.
Just recently, Mirendil published an article on the official web, entitled " Large-scale pre-training without a pre-training team ", which was sensational.
Three numbers appear below the title of the article:
One researcher for three and a half months at a cost of less than $10 million.
The money covered the token consumption of Agent, data processing and the total cost of research experiments. With these inputs, they really got a model of hundreds of billions of parameters: Mirendil-Base, a coded base model between 100 billion and 500 billion parameters.
In its article, Mirendil stated that over 99 per cent of the new model codes were not written by humans, but by Agent.
Human researchers are only responsible for determining direction, judging and stepping on brakes at key nodes. Writing codes, running experiments, miscalculation, etc. were all given to Agent.
In Mirendil ' s tests, some of the indicators of the new model have already approached mainstream first-line models such as Kimi K3, GLM-5.3 and DeepSeek-V4.
The age of "one-man-one-man-one-man-one-man-one-one-man-one-one-man-one-one-man-one-one-one-one-man-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-one-
01
According to official accounts, Mirendil-Base was the first product of the Mirendil series model, positioning was the code-specific base model, and the target user was a researcher and engineer inside the front-line AI laboratory.
In order to verify whether a person can make a model, Mirendil has put in place a number of harsh conditions. For example, there can be no ready-made pre-training framework and no inheritance code.
Researchers and Agent started from scratch and built everything for more than 100 days — training infrastructure, model architecture, training algorithms, data conduits — all written in their own right, except for some of the hot open source projects on the general libraries of PyTorch and JAX.
Of course, the word “self” here is usually about Agent.
Researchers are primarily responsible for deciding what to try, what to trust and intervening when needed. The specific actions of writing codes, running experiments and making errors in diagnosis are basically Agent running by himself.
In terms of data, Mirendil uses open data such as GitHub open source code, open web pages, open data sets, etc. The company emphasized that no line of data had been purchased during the training, that no distillation had taken place and that no synthetic data had been used.
In addition, Mirendil hopes to improve efficiency through better design. “We have achieved significant efficiency gains through improved structures, high-level reconciliations of data mix, and other improvements at the data level, as well as by the kernel optimization. This is what the announcement says.
The final product is Mirendil-Base, with a parameter of 100 billion-500 billion.
What's the level of the hundreds of billion parametric models that a man makes?
Mirendil compared five models, Kimi K3, GLM-5.3, Nemotron-550B-Base, DeepSeek-V4-Pro-Base, and Inkling, all large-parametric models of MoE structures.
In the cross-radon tests at the new code warehouse (the lower the score the more the model guesses the code, the better the guess), the results of the cross-classes of several models are as follows:
Kimi K3 ranks first, score 0.193; Mirendil-Base ranks second, score 0.222; followed by GLM-5.3, DeepSeek-V4, Nemotron.
In other words, on the partial coding dimensions, a pre-training model trained by a person and Agent team is already able to break hands with the finished product model of a large head plant.
It is important to note, however, that, as a rule, the matrix model performed better in cross-breeding tests than the later training model. Mirendil-Base, which is a base model, has a higher score than Kimi-K3, which shows that Kimi-K3 is much more powerful than the apparent score, and that Mirendil-Base has a greater difference than it looks.
According to the official presentation, Mirendil-Base has two more lines up: Large is between $500 billion and $2 trillion, and XL runs directly into 2 trillion and 10 trillion.
At present, Large is 80 per cent trained and XL is only 19 per cent in the beginning, but it's almost up to Large's current score.
That is to say, with currency capacity and Agent's grip, it's not a dream to train trillions of models — a threshold that many major AI factories have yet to touch.
02
The pre-training of hundreds of billions, if not hundreds of billions of parameters LLM, is extremely expensive and time-consuming, and a huge pile of people is required.
Some of the hundreds of researchers were looking at what models should look like and how to scaling; some managed the clean-up and matching of trillions of tokens; and others were holding thousands, if not hundreds of thousands of calculators, to ensure that the training never stopped.
Large AI plants can bring together enough talent, but financial and computing costs remain high.
The whole industry has been trying to downplay, and the central idea is to let AI train AI.
For example, Anthropic has come a long way.
It had previously disclosed that in May this year, Claude had written over 80 per cent of the consolidation codes in its code repository.
Google is doing similar things: AI programming AI agent AlphaEvolve has been used to optimize data centre movement, chip design and AI training processes.
However, the “AI Training AI” route is currently the only one that can make a difference.
Scientists, researchers, small teams, etc. wanted to train a model for themselves, which was almost impossible before.
Mirendil-Base has demonstrated that under certain conditions a person can also “hand-stamp” a model.
In an official web article, Mirendil stated that until recently, it had been a bottleneck in the pre-training of a competitive model: an expert team that took years to build.
Today, the situation is changing.
Once only a large research institution could do it, a researcher could do it now. When Agent took on most of the hands-on work, the bottlenecks shifted to arithmetic, data and the judgement of the person who commanded Agent.
"Everybody can train in models" has huge imagination.
Previously, most people and companies often rented a generic model and adapted it at the edge if they needed it.
Now, a chip design team, a robotic laboratory or a hospital can train a model for their own problems with their own data. The base model and the systems built around it will no longer be something that only a few large plants can make, but will become more widespread.
On the other hand, AI projects, which in the past required a large team to start, can now start with small teams or even one person; and more people will be able to try them as computing becomes cheaper and stronger. “One-person company” or “super individuals” are expected to become a reality.
The way people work will also change: as more work is given to Agent, the daily work of mankind will become less and less like “doing things hand in hand” and more and more like “leading a small laboratory”: setting the direction of research, designing experiments worth running, and determining what ideas are worth moving forward.
Mirendil paints a more spectacular vision: the next model will be built in part by Mirendil's own model, and every model that becomes stronger in building AI will make the next model cheaper and faster.
“We believe that the RSI has not reached full autonomy, but we are closer than ever. The company says.
In addition, the company has initiated post-training for Mirendil-Base. Mirendil-Base-Large and Mirendil-Base-XL at the front level are still in training.
03
Behind Mirendil are two former employees of Anthropic: Behnam Neyshabur and Harsh Mehta.
Neyshabur was born in Iran and went to university after 80. He went to the United States in 2011 and obtained a doctorate degree from the Chicago branch of Toyota Industrial University. After graduation, he worked in DeepMind for five years in Google, joined Anthropic at the end of 2024, led the Discovery team and studied how to involve AI as a scientist in research and development.
Mehta is Indian, graduated from Guwahati, Indian Institute of Technology, and worked with Neyshabur in DeepMind. With the addition of Anthropic, Mehta built the first version of the Anthropic internal automated AI research platform.
In December 2025, the two men left behind and set up Mirendil, and the company was positioned to help scientists build their own AI, not just direct AI-aided science.
In other words, they want to allow the scientific team to train itself and to use it in an iterative manner, without relying on the generic model of a large plant. The technical route behind it is the latest hot fire in the AI Circle.
In June this year, Mirendil announced $200 million in seed ship financing, with a post-investment valuation of $1 billion. The investors are a16z, Kaypeng Ying, Wei Da, etc.
The team is currently about 20 people, all researchers and engineers from Anthropic, XAI, Google DeepMind, Open AI, based in downtown San Francisco.
Mirendil-Base's appearance marked the first milestone for the company.
But then again, “one person trains a hundred billion parameter models” sounds good and the threshold is not much lower.
The cost of $10 million was nothing for large plants, and Mirendil, who had just received financing, could still afford it. But for the vast majority of individuals and small teams, it is still astronomical.
Not to mention, being able to direct Agent's work to such a great extent that he can't just pull an engineer. Mirendil has a high quality of talent and experience in the AI project and is in the first class throughout Silicon Valley.
It is true that the era of “one-man rubbing LLM” is near, but far from ordinary people.
Mirendil-Base ' s publish is the first sign of this long road. As to whether it can really lower the threshold of LLM, it also depends on the actual level of subsequent larger models and whether the methodology can be replicated on others.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.