How to Apply to Preferred Networks

Key Takeaways

Preferred Networks is Japan's most prominent deep learning company and one of the very small number of organizations globally that designs its own AI accelerators in-house alongside building foundation models, robotics platforms, and industrial AI applications.
PFN built its reputation on Chainer, the original define-by-run deep learning framework that influenced PyTorch, and continues to be a major open source contributor through Optuna, PFRL, ChainerCV, and PyTorch upstream, which makes a track record of meaningful OSS contribution a strong differentiator.
The company has deep, capital-backed partnerships with Toyota, FANUC, Mizuho, Hitachi, JXTG, ENEOS, and NTT, which means PFN engineers ship AI into real cars, factories, refineries, financial systems, and laboratories rather than only writing papers.
PFN designs the MN-Core family of custom AI accelerators with TSMC and operates its own MN-Core supercomputers for foundation model training, making it one of the rare places where a single engineer can touch silicon, compiler, framework, model, and application in the same building.
The PLaMo family of large language models is PFN's bet on enterprise-grade Japanese-language foundation models, and roles in pretraining, fine-tuning, evaluation, and deployment of PLaMo are a significant and growing share of open requisitions.
Interviews are rigorous and research-led, with deep paper or project walkthroughs, coding screens, ML and systems design rounds, and a one-hour talk for senior research candidates that is by far the highest-weighted single signal for those roles.
Resumes that quantify research and production impact (paper venues, benchmark improvements, latency, throughput, robot success rates, hardware utilization) and that match PFN's vocabulary score significantly higher in both reviewer triage and interview rapport.
Japanese language ability is helpful but not strictly required for many engineering and research roles, since core teams operate in English, while customer-facing and corporate roles do require business-level Japanese; PFN sponsors visas, supports relocation, and offers Japanese language training for international hires.
Compensation includes a competitive base salary in Japanese yen, a discretionary annual bonus, stock options under PFN's employee equity plan, comprehensive Japanese social insurance, and relocation support, with overall packages calibrated to be competitive with FAANG Tokyo offers for top research and engineering candidates.

About Preferred Networks

Preferred Networks, Inc. (PFN) is a Tokyo-headquartered Japanese deep learning and artificial intelligence company widely regarded as the country's most prominent and technically ambitious AI startup. Founded in March 2014 by Toru Nishikawa (CEO) and Daisuke Okanohara (Executive Vice President), PFN spun out of Preferred Infrastructure (PFI), the natural language processing and search company the same founders started in 2006 while still graduate students at the University of Tokyo. The company employs roughly 300 people, the overwhelming majority of whom are research scientists and engineers, and its headcount is intentionally kept small relative to the scope of its ambition. PFN is best known internationally as the original creator of Chainer, the define-by-run deep learning framework released in 2015 that pioneered the dynamic computation graph model later popularized by PyTorch, to which PFN officially migrated its research stack in 2019 in a widely cited gesture of intellectual honesty about where the broader ecosystem was heading. Strategically, PFN positions itself as a deep tech company that builds proprietary AI hardware, foundation models, and full-stack solutions for industrial applications rather than a consumer software company, and it has secured landmark partnerships and investments from Toyota Motor Corporation, FANUC, Mizuho, Hitachi, JXTG, ENEOS, NTT, and the Japan External Trade Organization, with cumulative funding exceeding 20 billion yen and a valuation that has historically made it Japan's most valuable AI unicorn. PFN's product portfolio spans autonomous driving research with Toyota, factory automation and robotics with FANUC, drug discovery and materials informatics through its joint venture Preferred Computational Chemistry, medical imaging through its subsidiary Preferred Networks Medical, the PLaMo family of large language models built specifically for Japanese-language enterprise use, the MN-Core family of custom AI accelerator chips designed in-house and manufactured with TSMC, and the Preferred Robotics consumer robot Kachaka. Culturally, PFN is research-led, technically intense, English-friendly within engineering teams, and genuinely international in its hiring, drawing talent from top universities and research labs across Japan, the United States, Europe, China, India, and Southeast Asia. Candidates evaluating PFN should expect a company that takes long-horizon technical bets seriously, that values published research and open source contributions alongside production engineering, and that has the rare combination in Japan of a deep research culture, real industrial customers, and the engineering autonomy to design its own silicon.

Application Process

1
Search and apply through preferred
Search and apply through preferred.jp/en/careers, which lists open positions across research, engineering, robotics, hardware, business, and corporate functions; most technical postings are dual-listed in Japanese and English, and you can apply in either language without penalty if the role description does not require business-level Japanese.
2
Submit a resume and a substantive cover letter or motivation statement; for rese
Submit a resume and a substantive cover letter or motivation statement; for research and engineering roles, also include a publications list, GitHub or portfolio links, and any conference talks or open source maintainership relevant to the team you are applying to, since reviewers explicitly weigh public technical evidence.
3
Expect a recruiter or talent partner screen within one to three weeks of applyin
Expect a recruiter or talent partner screen within one to three weeks of applying for shortlisted candidates; the recruiter will calibrate on language ability (Japanese, English, or both), visa and relocation requirements, motivation for PFN versus competing offers from FAANG Tokyo, OpenAI, DeepMind, or domestic alternatives, and target team within research, engineering, robotics, hardware, or applications.
4
Technical screens follow and typically include a 60 to 90 minute coding or resea
Technical screens follow and typically include a 60 to 90 minute coding or research conversation conducted in English or Japanese depending on candidate preference; software engineering screens cover algorithms, data structures, and systems fundamentals, while research screens focus on a deep walkthrough of one of your published papers or a major project, with the interviewer probing your contribution, the methodology, the failure modes, and the open questions you would pursue next.
5
Onsite or virtual loops typically include four to six interviews covering coding
Onsite or virtual loops typically include four to six interviews covering coding, machine learning depth, system design or research design, a hiring manager conversation, a behavioral and culture-fit round, and a cross-functional collaboration round with a partner team such as robotics, MN-Core hardware, or one of the industrial application groups.
6
Senior, staff, principal, and research scientist candidates often present a one-
Senior, staff, principal, and research scientist candidates often present a one-hour talk on a major past project or research direction to a panel of PFN scientists and engineers, followed by extensive Q and A; this talk is the single highest-weighted signal for senior research hires and should be prepared with the same rigor as a top-tier conference presentation.
7
Offers are typically extended within two to four weeks of the final loop and inc
Offers are typically extended within two to four weeks of the final loop and include base salary in Japanese yen, a discretionary annual bonus, stock options under PFN's employee equity plan, comprehensive Japanese social insurance and benefits, relocation support for international hires including visa sponsorship under the Highly Skilled Foreign Professional or Engineer/Specialist in Humanities/International Services categories, and Japanese language training subsidies for non-Japanese-speaking employees who want to invest in long-term tenure.

Resume Tips for Preferred Networks

recommended

Lead with measurable research and engineering impact rather than responsibilitie

Lead with measurable research and engineering impact rather than responsibilities: cite paper acceptances at top venues (NeurIPS, ICML, ICLR, CVPR, ICRA, RSS, ACL, EMNLP, ISSCC, MICRO), benchmark improvements with the baseline you started from, production model latency or throughput gains, hardware utilization improvements, or robot task success rate increases with concrete numbers and the evaluation protocol used.

recommended

Show deep learning and ML systems depth explicitly

Show deep learning and ML systems depth explicitly. PFN cares about PyTorch internals, distributed training (data, model, pipeline, and tensor parallelism), foundation model pretraining and fine-tuning, reinforcement learning, simulation-to-real transfer, computer vision, robotics manipulation and locomotion, large language models, materials and molecular informatics, and medical imaging, and these terms should appear naturally in your bullets when they apply.

recommended

Surface hardware-software co-design experience candidly if you have it

Surface hardware-software co-design experience candidly if you have it. PFN designs its own MN-Core accelerators in-house, so candidates with experience in CUDA, ROCm, MLIR, TVM, custom kernels, compiler backends, RTL design, high-level synthesis, or ASIC verification are highly differentiated and should highlight that work prominently rather than burying it.

recommended

Translate industrial and applied AI experience into PFN-relevant language

Translate industrial and applied AI experience into PFN-relevant language. If you have worked on autonomous driving, factory automation, drug discovery, materials science, medical AI, or industrial robotics, name the customer, the deployment scale, the constraints (latency, safety, certification, regulation), and the engineering choices you made, since PFN's business model is built on shipping AI into hard physical systems.

recommended

List languages and frameworks honestly with depth indicators

List languages and frameworks honestly with depth indicators. PFN writes meaningful production code in Python, C++, Rust, CUDA, and Go, runs PyTorch and JAX as primary research stacks, uses Kubernetes and Slurm for cluster orchestration, and ships robotics code in ROS and ROS 2; padding the list with twenty technologies you used once is a negative signal for engineers and scientists who will read the resume.

recommended

Include open source and community contributions

Include open source and community contributions. PFN is a major open source contributor through Chainer (legacy), Optuna, PFRL, PyTorch contributions, ChainerCV, and many smaller libraries, and a track record of meaningful PRs to mainstream OSS, sustained maintainership of a useful library, or significant contributions to academic codebases is a strong positive signal.

recommended

Indicate Japanese language ability honestly and precisely using the JLPT scale (

Indicate Japanese language ability honestly and precisely using the JLPT scale (N5 through N1) or descriptive levels (none, conversational, business, native), since the role's required language depends on team composition and customer interface; many engineering and research teams operate primarily in English, but partnerships with Toyota, FANUC, and other Japanese industrial customers are conducted in Japanese.

recommended

Keep the resume to one or two pages with a clean, conservative layout and consis

Keep the resume to one or two pages with a clean, conservative layout and consistent typography in either English or Japanese, and submit as a PDF; PFN reviewers read across many language backgrounds and a dense but legible single page in clear English or natural Japanese outperforms a five-page narrative every time.

ATS System: In-house Application System

Preferred Networks accepts applications directly through its careers site at preferred.jp/en/careers (English) and preferred.jp/careers (Japanese), backed by an in-house application workflow rather than a major commercial ATS such as Workday, Greenhouse, or Lever. Each posting routes to a structured application form that collects your resume, cover letter or motivation statement, contact information, work authorization status, language abilities (Japanese and English), and links to publications, GitHub, portfolio, or other public technical evidence. Applications are reviewed by PFN talent partners and hiring managers in collaboration with the technical teams, and status updates are communicated by email rather than through a self-serve candidate portal.

S
u
b
m
i
t
a
c
l
e
a
n
P
D
F
r
e
s
u
m
e
i
n
e
i
t
h
e
r
E
n
g
l
i
s
h
o
r
J
a
p
a
n
e
s
e
w
i
t
h
s
t
a
n
d
a
r
d
s
e
c
t
i
o
n
h
e
a
d
i
n
g
s
(
E
x
p
e
r
i
e
n
c
e
,
E
d
u
c
a
t
i
o
n
,
P
u
b
l
i
c
a
t
i
o
n
s
,
S
k
i
l
l
s
,
O
p
e
n
S
o
u
r
c
e
)
s
o
r
e
v
i
e
w
e
r
s
c
a
n
s
c
a
n
i
t
q
u
i
c
k
l
y
;
a
v
o
i
d
t
a
b
l
e
s
,
m
u
l
t
i
-
c
o
l
u
m
n
l
a
y
o
u
t
s
,
h
e
a
d
e
r
s
a
n
d
f
o
o
t
e
r
s
,
a
n
d
e
m
b
e
d
d
e
d
i
m
a
g
e
s
,
a
l
l
o
f
w
h
i
c
h
c
o
r
r
u
p
t
p
a
r
s
i
n
g
.
M
i
r
r
o
r
e
x
a
c
t
t
e
r
m
i
n
o
l
o
g
y
f
r
o
m
t
h
e
j
o
b
d
e
s
c
r
i
p
t
i
o
n
(
f
o
r
e
x
a
m
p
l
e
u
s
e
'
P
y
T
o
r
c
h
'
,
'
d
i
s
t
r
i
b
u
t
e
d
t
r
a
i
n
i
n
g
'
,
'
f
o
u
n
d
a
t
i
o
n
m
o
d
e
l
'
,
'
r
e
i
n
f
o
r
c
e
m
e
n
t
l
e
a
r
n
i
n
g
'
,
'
s
i
m
u
l
a
t
i
o
n
t
o
r
e
a
l
'
,
'
C
U
D
A
'
,
'
M
N
-
C
o
r
e
'
,
'
R
O
S
'
,
'
P
L
a
M
o
'
,
o
r
s
p
e
c
i
f
i
c
p
r
o
d
u
c
t
o
r
p
r
o
j
e
c
t
n
a
m
e
s
)
s
o
k
e
y
w
o
r
d
m
a
t
c
h
i
n
g
s
c
o
r
e
s
y
o
u
c
o
r
r
e
c
t
l
y
.
I
n
c
l
u
d
e
d
i
r
e
c
t
l
i
n
k
s
t
o
y
o
u
r
m
o
s
t
i
m
p
o
r
t
a
n
t
p
a
p
e
r
,
y
o
u
r
G
i
t
H
u
b
p
r
o
f
i
l
e
,
a
n
d
o
n
e
o
r
t
w
o
s
p
e
c
i
f
i
c
r
e
p
o
s
i
t
o
r
i
e
s
o
r
P
R
s
t
h
a
t
d
e
m
o
n
s
t
r
a
t
e
t
h
e
d
e
p
t
h
c
l
a
i
m
e
d
i
n
y
o
u
r
r
e
s
u
m
e
.
W
r
i
t
e
a
s
u
b
s
t
a
n
t
i
v
e
c
o
v
e
r
l
e
t
t
e
r
o
r
m
o
t
i
v
a
t
i
o
n
s
t
a
t
e
m
e
n
t
(
2
5
0
t
o
5
0
0
w
o
r
d
s
)
t
h
a
t
n
a
m
e
s
t
h
e
s
p
e
c
i
f
i
c
t
e
a
m
o
r
p
r
o
j
e
c
t
a
t
P
F
N
y
o
u
w
a
n
t
t
o
j
o
i
n
a
n
d
e
x
p
l
a
i
n
s
w
h
y
y
o
u
r
b
a
c
k
g
r
o
u
n
d
f
i
t
s
,
s
i
n
c
e
g
e
n
e
r
i
c
a
p
p
l
i
c
a
t
i
o
n
s
a
r
e
f
i
l
t
e
r
e
d
o
u
t
q
u
i
c
k
l
y
.
A
p
p
l
y
w
i
t
h
i
n
t
h
e
f
i
r
s
t
t
w
o
w
e
e
k
s
o
f
a
p
o
s
t
i
n
g
g
o
i
n
g
l
i
v
e
,
s
i
n
c
e
r
e
v
i
e
w
e
r
a
t
t
e
n
t
i
o
n
i
s
h
i
g
h
e
s
t
e
a
r
l
y
i
n
t
h
e
c
y
c
l
e
.

Complete In-house Application System Resume Guide →

Interview Culture

PFN interviews are deliberately rigorous, intellectually deep, and grounded in real research and production scenarios rather than puzzle questions detached from the work.

Expect interviewers to be working scientists, engineers, or robotics researchers who currently lead or contribute to the projects you would be joining, and expect them to push past the first answer until they understand exactly how you think, where the boundaries of your knowledge actually are, and what you would do with more time, more compute, or more data. A common opening is a behavioral conversation that quickly turns into a project or paper walkthrough, and your interviewer will keep asking why, what would break if, and what you would try next at every layer until the methodology, the trade-offs, and the failure modes are fully exposed. For research candidates, the paper walkthrough and the one-hour talk are by far the highest-weighted signals; PFN scientists evaluate originality, technical depth, ability to articulate contribution versus prior work, statistical and methodological rigor, and how you handle pointed questions about limitations or alternative interpretations of your results. Coding rounds for engineering tracks tend to focus on algorithms, data structures, numerical computation, and systems concurrency, with an emphasis on correctness, edge cases, and clear reasoning rather than speed-run optimization; for ML systems and infrastructure tracks, you will frequently be asked to design or debug distributed training pipelines, reason about GPU memory hierarchies, or sketch a custom CUDA kernel rather than solve pure algorithm puzzles. System design and ML design rounds are unusually concrete because PFN actually trains foundation models on its own MN-Core supercomputers and ships robots into customer factories, so expect prompts about designing a distributed training stack for a 100B-parameter model, a real-time perception pipeline for a manipulator, a safety architecture for a consumer robot, an active learning loop for an industrial inspection task, or a simulation-to-real pipeline for a legged robot. For hardware candidates, expect deep questions on dataflow architectures, sparsity and quantization, memory bandwidth, compiler IRs, and the trade-offs between systolic arrays and SIMT execution. Behaviorally, PFN screens hard for intellectual honesty, technical curiosity, ownership, and the ability to collaborate across very different disciplines (a robotics engineer must talk to a chip designer must talk to an LLM researcher), and interviewers respond well to candidates who admit what they do not know, describe failures with what they learned, and demonstrate genuine excitement about long-horizon problems. They respond poorly to confident bluffing, hype-driven answers, or contempt for either academic rigor or production constraints. The tone is professional, calm, and methodical rather than performative, conducted in English or Japanese based on candidate preference and team composition, and decisions are made by full loop debriefs where every interviewer must justify their recommendation with concrete evidence from the conversation.

What Preferred Networks Looks For

Researchers and engineers with genuine depth in deep learning, robotics, ML systems, computer vision, NLP, reinforcement learning, materials and molecular informatics, medical AI, or AI hardware, demonstrated through publications, open source, or shipped systems rather than buzzwords.
Hands-on builders who have shipped and operated real ML systems with real customers, real latency budgets, real safety constraints, or real failure consequences, especially in industrial, robotics, healthcare, or scientific computing contexts.
Scientists who can frame a research problem clearly, design rigorous experiments, write the code themselves, interpret negative results honestly, and translate research insights into production systems that customers actually use.
Pragmatic generalists who can move fluidly between PyTorch model code, distributed training infrastructure, custom CUDA or MN-Core kernels, simulation environments, and robotics middleware without retreating into a single discipline.
Strong written and verbal communicators in English, Japanese, or both, since PFN engineering operates internationally while customer-facing work with Toyota, FANUC, ENEOS, and other partners is conducted in Japanese and bridges between the two are constantly needed.
Owners who take end-to-end responsibility, including the unglamorous work of dataset curation, labeling pipelines, evaluation harnesses, regression testing, and documentation of model behavior in edge cases.
Collaborative teammates who are comfortable in a small, dense organization where a robotics researcher might pair with a chip designer in the morning and an LLM researcher in the afternoon, and who respect cross-disciplinary expertise rather than dismissing it.
Candidates with a long-term mindset who are excited by the multi-year arc of building Japanese-grounded foundation models, advancing legged and manipulator robotics, scaling MN-Core silicon, and applying AI to drug discovery and materials science problems that take years to mature.

Frequently Asked Questions

What ATS does Preferred Networks use, and how should I apply?

Preferred Networks accepts applications through its careers site at preferred.jp/en/careers (English) and preferred.jp/careers (Japanese), backed by an in-house applicant tracking workflow rather than a major commercial ATS like Workday or Greenhouse. Submit a single clean PDF resume in either English or Japanese along with a substantive cover letter or motivation statement, and include publications, GitHub, and portfolio links inline. Apply directly through the official careers page rather than via third-party aggregators, and avoid creating duplicate applications across multiple roles in the same week.

Do I need to speak Japanese to work at Preferred Networks?

It depends on the role. Many research, engineering, robotics, and hardware teams operate primarily in English internally, and PFN actively recruits international talent under Japan's Highly Skilled Foreign Professional and Engineer visa categories, so business-level Japanese is not required for those roles. However, customer-facing positions with Toyota, FANUC, ENEOS, and other Japanese industrial partners, as well as most corporate and business roles, do require business-level Japanese (typically JLPT N2 or higher). Confirm the language expectation for the specific requisition with your recruiter during the initial screen.

Does PFN sponsor visas and relocation for international hires?

Yes. PFN routinely sponsors Japanese work visas for qualifying technical and research roles, most commonly under the Highly Skilled Foreign Professional category which provides accelerated permanent residency eligibility, or the Engineer/Specialist in Humanities/International Services category. PFN also offers relocation support including flights, temporary accommodation, shipping, and assistance with the housing and bank account setup that is notoriously difficult in Japan, and many teams subsidize Japanese language lessons for employees who want to invest in long-term tenure.

What is the interview process for research scientist roles?

A typical research scientist loop includes a recruiter screen, an initial paper or project deep-dive interview with a PFN scientist, a coding screen, a hiring manager conversation, an onsite or virtual loop of four to six rounds covering ML depth, research design, coding, and behavioral fit, and a one-hour talk by the candidate to a panel of scientists and engineers on a major past project or research direction. The talk is the single highest-weighted signal for senior research hires and should be prepared with the same rigor as a top-tier conference presentation.

What is the interview process for software and ML engineering roles?

A typical engineering loop includes a recruiter screen, a 60 to 90 minute technical screen with a hiring engineer focused on coding, algorithms, and systems fundamentals, a hiring manager conversation, and an onsite or virtual loop of four to six rounds covering coding, ML or systems design, architecture deep dive, behavioral fit, and a cross-functional collaboration round with a partner team such as robotics, MN-Core hardware, or one of the industrial application groups. Senior and staff candidates often add a project deep-dive presentation.

How should I prepare for PFN ML and systems design interviews?

PFN design rounds are unusually concrete because the company actually trains foundation models on its own MN-Core supercomputers and ships robots into customer factories. Prepare by studying distributed training (data, model, pipeline, and tensor parallelism), foundation model pretraining and fine-tuning, GPU memory hierarchies, custom kernels, real-time perception and control for robotics, simulation-to-real transfer, active learning, and safety architectures for physical systems. Be ready to defend specific trade-offs, name the failure modes, and explain how you would evaluate the system in production rather than recite reference architectures.

What technical skills matter most across PFN roles?

Across most research and engineering tracks, the highest-leverage skills are deep PyTorch and JAX proficiency, distributed training experience, strong Python and C++ engineering, CUDA or accelerator programming for ML systems and hardware tracks, ROS and robotics middleware for robotics tracks, foundation model pretraining and fine-tuning for LLM tracks, and rigorous experimental methodology across the board. A track record of top-venue publications, meaningful open source contributions, or shipped production ML systems is a strong differentiator regardless of track.

What is compensation and equity like at Preferred Networks?

PFN compensation includes a competitive base salary calibrated to level and role and denominated in Japanese yen, a discretionary annual bonus typically expressed as a number of months of base, stock options under the PFN employee equity plan, comprehensive Japanese social insurance covering health, pension, and unemployment, commuting allowance, and relocation support for international hires. Overall packages for senior research and engineering candidates are calibrated to be competitive with FAANG Tokyo and OpenAI Tokyo offers, though headline numbers in Japanese yen often look lower than US-denominated equivalents until you account for cost of living and tax differences.

What is the working culture like day to day?

PFN is research-led, technically intense, and unusually international for a Japanese company, with hiring across Japan, the United States, Europe, China, India, and Southeast Asia. Engineering teams typically operate in English internally with Japanese as the bridge language for customer interactions, the office in Otemachi Tokyo is the primary hub with hybrid work supported, and the culture values published research and open source contribution alongside production engineering. Long-tenured employees are common, internal mobility between research, engineering, robotics, and hardware tracks is genuinely supported, and the small headcount (around 300) means individual engineers and scientists have unusually high leverage on the company's direction.

How do I differentiate myself if I do not have prior experience with foundation models or robotics?

You do not need prior foundation model or robotics experience to succeed at PFN. What matters is demonstrated depth in adjacent areas (any rigorous ML research, distributed systems, scientific computing, computer graphics, computer vision, NLP, control theory, embedded systems, compilers, or chip design) and the ability to reason rigorously about trade-offs in your interviews. Translate your past work into the language of accuracy, latency, throughput, sample efficiency, safety, and customer impact, contribute meaningfully to open source if you can, and study PFN's recent papers, MN-Core technical disclosures, and PLaMo announcements enough to ask informed questions about how the role connects to the company's technical strategy.

Open Positions

Preferred Networks currently has 3 open positions.

Check Your Resume Before Applying → View 3 open positions at Preferred Networks

How to Apply to Preferred Networks

Key Takeaways

About Preferred Networks

Application Process

Resume Tips for Preferred Networks

Lead with measurable research and engineering impact rather than responsibilitie

Show deep learning and ML systems depth explicitly

Surface hardware-software co-design experience candidly if you have it

Translate industrial and applied AI experience into PFN-relevant language

List languages and frameworks honestly with depth indicators

Include open source and community contributions

Indicate Japanese language ability honestly and precisely using the JLPT scale (

Keep the resume to one or two pages with a clean, conservative layout and consis

ATS System: In-house Application System

Interview Culture

What Preferred Networks Looks For

Frequently Asked Questions

Open Positions

Related Resources

Similar Companies

Sources