In the autumn campus recruitment for large model related roles, those fresh graduates from top prestigious universities who have landed zero offers

"You're being too optimistic. I've been preparing since July and still haven't landed a single offer," Ximu, a fresh graduate from a Double 211 university, replied dejectedly when asked about his results in this year's autumn recruitment.

He has completed internships at leading semiconductor firms and internet companies, and most of his internship roles were related to large model inference and training framework optimization. After failing to secure any interviews for Agent positions, he decided to abandon the lower-threshold Agent development to boost his competitiveness in autumn recruitment, and shifted to training Infrastructure which has higher technical barriers and requires hands-on access to GPU clusters.

He once almost got an offer from ByteDance, and advanced to the fourth round of interview before getting no further feedback. By September, he found the situation even worse, with far fewer interview opportunities available.

He is not the only one facing such a dilemma. Last year, almost no one in his postgraduate lab landed internships at major tech firms, but all of them eventually secured proper positions upon graduation. This year, everyone in the lab has internship experience, and some even have three internships under their belt. But as Ximu observed by October, no one has received a formal full-time offer yet.

A student who completed his undergraduate study at one of the top 5 universities in China and is pursuing a direct PhD program at a top 2 university, with foundation model internship experience, has attended interviews for model teams including Seed, Tencent Hunyuan and Alibaba, but has not received any clear offer result for a long time.

Another master's student with 5 A-class papers (as first author or co-first author) and more than 600 total paper citations started applying for large model talent programs of various companies in July. The vast majority of his application processes got stuck after the first interview and never received any follow-up responses.

Young talents that model companies have been competing fiercely for over the past two years have suddenly lost their appeal in the 2026 autumn recruitment season.

A headhunter who has long served AI-native companies told us that many model companies were building their teams last year, and they were scrambling for talents who could independently push a specific technical direction forward. This year, their teams are already established, and they no longer need fresh graduates with similar capabilities. In other words, they hope new joiners can solve problems that the existing team has not yet managed to tackle.

The recruitment structure of ByteDance is quite representative. The group's 2026 spring ByteIntern recruitment plan covers more than 7,000 positions, of which R&D roles exceed 4,800. Meanwhile, Seed launched a separate global talent special program of around 100 people on April 1, there will be no separate Top Seed program for 2027 graduates for Seed recruitment, and the separate Jindouyun talent program for other cutting-edge talent recruitment has also been cancelled.

The tide has shifted, and fresh graduates in this industry are not welcoming a "warm winter".

01

Large model companies no longer need the same type of talents

Jing Jing studies at Fudan University. At the end of 2025, as a researcher focusing on post-training algorithm research, he had three excellent internship offers at the same time: Agent engineer positions from top internet giants and leading foundation model vendors, as well as a foundation model R&D position from another leading internet company.

Choosing one out of three was not a difficult decision. Inside the large model industry, there has long been an unwritten ranking for technical positions: pre-training is closest to the source of model capability, followed by post-training, and application and engineering development rank lower.

Although this ranking does not necessarily reflect the actual technical contribution, for students who want to engage in cutting-edge R&D, joining a foundation model team usually means better resources, larger training scale and broader career options.

He chose the foundation model R&D position without hesitation. But after onboarding, the actual experience was far from what he had imagined.

In the first week of his employment, his main tasks were to get familiar with documents and configure the development environment. For a long period after that, he was only assigned tasks including code maintenance, testing and research. Nearly three months passed, and he had not accumulated much practical output. It was not until April that he was formally assigned to work related to pre-training and post-training, allocated with resources to build the training workflow. But after the workflow was built, his tasks turned to maintaining the stable operation of the system again.

It was not until August that he was informed that he could not get a full-time conversion from the internship. His autumn recruitment process was delayed, and he still has not landed any offer to date.

Jing Jing's experience may have personal reasons, but it also points to an easily overlooked problem: joining a foundation model team is not the same as actually participating in core model R&D.

In the past, model companies built their teams from scratch, and needed a large number of researchers and engineers to complete basic work such as the training framework, data processing and experimental process. A foundation model internship used to mean access to important R&D tasks. But as the model training system gradually matures, work starts to be stratified. A small number of core researchers determine the training strategy and technical route, engineers are responsible for the development and optimization of different modules, and more work turns into maintaining the already running systems.

A headhunter used video models as an example. In 2026, Seedance has already proved ByteDance's technical capability in video generation. But a department that already has a mature R&D team and has completed phased technological breakthroughs may no longer need to recruit a large number of fresh graduates with similar experience at a high premium.

"A team that has already built the world's best technology will no longer spend a high premium to recruit people who cannot bring additional value to the team," the headhunter said.

There is also a deeper constraint: computing power. For some cutting-edge model teams, the current bottleneck comes more from computing power rather than the number of R&D personnel.

A fresh graduate who joined a leading model vendor this year and works on training Infrastructure said that for large model training, doubling the number of engineers does not mean you can run twice as many experiments at the same time. If the GPU cluster scale, experiment scheduling and training resources are already close to the upper limit, more researchers may not speed up model R&D in the same proportion.

The shortage of computing power may in turn increase the value of talents specialized in training system optimization, parallel computing and algorithm efficiency, and also indicate a talent logic different from that of the traditional internet industry. Now that model companies already have sufficient personnel and mature training systems, the R&D increment brought by continued expansion has become a problem that needs careful calculation.

The most cutting-edge overseas model companies have already seen changes in their talent structure. On October 16, Anthropic released a research on the changes of its employment structure. At the beginning of 2025, junior technical personnel accounted for 27% of new recruits in the company's technical department, and this proportion dropped to 9% in 2026.

In January this year, Anthropic had already started internal discussions on the declining demand for junior technical personnel. As Claude can complete more work such as code writing and debugging, some teams have lowered the recruitment priority for junior engineers. The report shows that by April, about one third of machine learning managers and one fifth of software engineering managers said that AI had led them to reduce recruitment.

It is worth noting that Anthropic's technical team is still expanding, and the absolute number of recruits has not stopped growing. What has changed is the position of employees with different qualifications and capabilities in the recruitment process. The growth of leading Silicon Valley tech companies like Anthropic has begun to no longer rely entirely on headcount expansion to increase output.

02

The trend shifts too fast, and experience may depreciate

The change in recruitment quotas of model companies is not the whole story. The fickleness of technology trends makes it even harder for fresh graduates to adapt.

Ximu chose training Infrastructure precisely to avoid such changes.

He used to work on Agent development. When looking for internships in March, he mainly applied for Agent-related positions. But soon he realized that there was a problem with Agent development: the entry threshold was getting lower and lower.

In his understanding, a student who wants to accumulate Agent development experience only needs a computer to call models, connect tools and build his own projects. In contrast, training Infrastructure requires real GPU resources and large-scale system practice, which ordinary students can hardly get equal access to at school.

The lab he works in happens to have a GPU cluster that allows him to work on training system optimization. So he decided to shift to Infrastructure.

This strategy did help him rule out some competitors, but also sent him into an "arena" where top talents are more concentrated.

When interning in the Infrastructure position of a major tech company, he found that among the young full-time employees around him, people with Double 985 academic backgrounds are very common, and there are also many masters and doctors from top universities. Compared with these people, his academic background has no advantage, and the training scale he can get access to is not large enough.

What is more awkward is that some smaller model training teams even think his experience does not match the actual business.

Ximu mentioned that he once interviewed for an autonomous driving model team. The team mainly trains models with a parameter scale of about 1 billion, while the training system he had access to before is of larger scale. As a result, this experience that he originally regarded as an advantage failed to help him get the position, and the recruiter replied: "The training scale you were exposed to during your internship is larger than ours, so your experience does not match our needs."

He is worried that he will eventually fall into the dilemma of "unable to land a good position while being overqualified for lower-level roles".

A track with higher technical barriers, while filtering out ordinary candidates, also means that the remaining competitors are usually stronger.

Similar changes have also taken place in the multimodal field.

This year, video generation models continue to be upgraded, and the technical value of multimodality has been repeatedly verified, but this does not seem to have directly translated into employment opportunities for fresh graduates during the autumn recruitment season.

The sense of mismatch is very common. Some people originally targeted the video model teams of leading tech giants, but later shifted their goals to 3D generation, world model or embodied intelligence companies.

A headhunter confirmed that he has also met candidates who shifted from multimodal background to embodied intelligence. The reason is simple: positions for world model, 3D generation and embodied intelligence have overlapping requirements for capabilities such as vision, spatial representation and video prediction. As embodied intelligence companies are continuously established, new teams need to absorb talents from adjacent fields.

But new demands do not mean that high premiums will always exist.

According to the headhunter's observation, some embodied intelligence companies were in a hurry to complete their R&D teams last spring. By autumn, the recruitment requirements have become more specific: companies no longer only need people familiar with multimodality or 3D generation, but also need candidates who can solve problems in the next stage.

Shifting from one popular direction to another cannot guarantee that you will always get high salary premiums.

This batch of fresh graduates happen to believe in a principle: to go to the top foundation model teams to train foundation models. The foundation model tracks are divided into first-tier, quasi-first-tier, second and third-tier, and a first-tier foundation model position is regarded as a "doomsday ticket".

"The large model industry is a winner-takes-all game. If you can't get into a first-tier model team, your career in this field will not go smoothly," Xie Xin told us. The cruelest part of autumn recruitment is the process where people hold such perceptions and standards, but are forced to gradually lower their expectations.

Last year, her senior fellow apprentice still had the opportunity to participate in core training. This year, Xie Xin successfully got a full-time conversion from internship, but only through extreme hard work. Working overtime on weekends is the norm. If an experiment crashes, she has to spend several times more time to debug, even if 95% of her time is spent cleaning data, and she doubts that what she is doing is meaningless.

The 2025 AI Index from Stanford provides an example of the rapid popularization of technical capabilities: from November 2022 to October 2024, for models that reach the GPT-3.5 level MMLU performance, the inference price per million tokens dropped from about 20 dollars to 0.07 dollars, a decrease of more than 280 times.

The drop in inference cost is not equivalent to a drop in R&D talent salary. But it shows that model capabilities that used to be expensive and scarce can become cheap and widespread in a very short time.

In June this year, PwC analyzed more than 1 billion recruitment advertisements worldwide and found that from 2019 to 2025, the job skill composition of the occupations most affected by AI changed 2.2 times more than that of the least affected occupations.

For practitioners in the large model industry, when a technology is still in the exploration stage, few people master it, and the experience itself has value. When it gradually forms mature tools, open source frameworks or standard engineering processes, for young people who rely on mastering these technologies, it means their career capital depreciates in advance.

03

The dream-making field in the AI era: the million-dollar annual salary has not disappeared

The situation presented in the talent market is more complicated.

In September, a report released by Maimai showed that ByteDance is still the company with the largest number of new AI positions released according to its statistical caliber, and the number of new AI positions released by DeepSeek increased by 601.40% year-on-year. In the global market counted by PwC, positions requiring AI skills still have an average salary premium of 62%, which is further higher than that of last year. The high-salary era of the AI industry has not ended yet, and it is still a dream-making field.

However, the price of different types of "talents" has begun to diverge.

A headhunter who has long served AI companies told us that in 2025, some model companies were willing to offer high salaries to outstanding fresh doctoral graduates in order to form their teams as soon as possible. The competition among model vendors also boosted this trend, which reached a small peak last year.

Even this year, for candidates who have real multiple options, how much salary they can get from a model company is not the only problem. They usually graduate from top 2 universities, hold core papers and rich internship experience, and they are the group with the "smallest shadow area" in this year's autumn recruitment. How much training resources can they get access to? Can they participate in important experiments? Can they influence the next iteration of the model? These are the bases for them to evaluate offers.

But this is only the trouble of a few people.

For more young people who want to join foundation model teams, the first problem to solve is whether they can get an offer at all.

In the increasingly crowded job market for large model positions, academic background has become a prominent screening criterion again.

Ximu believes that the biggest difficulty in this year's job hunting is not that he cannot answer technical interview questions, but that his resume cannot even pass the initial screening to get an interview. He was even commented in person by a colleague who had already joined a major tech company that his academic background was not good enough.

Some unwritten rules unknown to outsiders are quietly spread inside large model companies. A leading internet company has a list of 33 universities around the world, and people whose academic background is not on the list cannot even get access to interviews; other major tech companies have special preference for fresh graduates from a small number of universities, but even if these people get in, most of them are arranged in marginal positions.

But academic background cannot explain all the problems. Among the group of fresh graduates participating in autumn recruitment that we communicated with, candidates with top 2 doctoral background, multiple top conference papers and multiple foundation model internship experiences are also experiencing long waiting periods.

This shows that the improvement of screening threshold does not simply mean that academic background replaces R&D capability, but that enterprises have more conditions to choose from. They not only require good university background and excellent papers, but also good internship teams and in-depth training experience, and it is best to prove that they have already achieved results in an important technical problem. Some talent programs that were originally targeted at a small number of top candidates may become even harder to enter this year.

"Last year, I might just not be able to get the best offer, but this year, I may not even be able to secure a position in the talent program."

One interviewee raised a question: how excellent do you need to be now to get a foundation model offer from BAT?

His question is real, but it is destined to be difficult to get an answer.

Another fresh graduate, who has already got a full-time conversion from internship at a leading internet company, still