Liang Wenfeng and Yang Zhilin Start Answering Different Questions

Produced by Huxiu Technology Group

Author: Song Sihang

Editor: Miao Zhengqing

Header image: Visual China

Moonshot AI, hereinafter “Kimi,” and DeepSeek have become the two most closely watched companies in the large-model sector in 2026. At the same time, the two companies are looking increasingly alike.

The first similarity is financing. Over the past six months, Kimi has completed several funding rounds in succession, with its valuation rising quickly. DeepSeek, long seen as a company that was not short of money, has also finally started fundraising. Two years ago, that would have been hard to imagine.

At the time, the two companies represented two very different paths in the large-model industry.

Kimi was the star startup, raising money frequently and expanding steadily. DeepSeek, by contrast, looked unusual. Liang Wenfeng rarely appeared in fundraising headlines, and the outside world’s strongest impression of DeepSeek was always its technology.

But this year, one change has become increasingly clear: both companies are now sitting at the capital market table.

The second similarity is their technical road map. For a long time, outsiders liked to compare the two companies, but in practice, they were not doing exactly the same thing.

Kimi has always been closer to products. From long context and search to later work on agents and coding, Moonshot AI has kept trying to turn model capabilities into products.

DeepSeek, by contrast, has looked more like a research institution. Whether with V2, R1 or V3, industry discussion has centered on the model capabilities themselves. But starting with this year’s V4, DeepSeek has also shifted its technical focus toward long context, coding and agents.

The boundaries between the two companies are becoming blurred.

Kimi is putting more emphasis on foundational model capabilities, while DeepSeek is paying more attention to inference efficiency, engineering systems and large-scale deployment. Even their hiring priorities are beginning to overlap.

Huxiu has noted that both Kimi and DeepSeek have recently continued hiring for harness-related roles. For large-model companies, a harness is not the model itself, but the infrastructure behind it. Training, inference, scheduling and resource management all depend on this system.

In other words, when both companies expand their harness teams at the same time, it shows they are no longer focused only on model capability. They are also focused on how models can be trained faster, run more reliably and be used at lower cost.

Beyond financing and technical direction, the founders’ management styles also share some similarities.

Yang Zhilin and Liang Wenfeng are both classic technical founders. Neither likes standing in the spotlight, and both rarely volunteer a narrative. Compared with markets, marketing or fundraising, they would rather talk about models, algorithms and technology itself.

Over the past few years, “technological idealism” has become perhaps the most consistent label outsiders have attached to both companies.

But after fundraising, the two companies may again head toward a new fork in the road.

Where Will DeepSeek Spend Its Money?

For DeepSeek, the biggest change this year is fundraising. For a long time, the outside world’s impression of DeepSeek was that it “didn’t lack money.”

High-Flyer Quant, which stands behind Liang Wenfeng, has the capacity to keep investing in AI research and development. Compared with many large-model companies that need constant fundraising to maintain their training cadence, DeepSeek has always looked somewhat unusual.

That is why, when news of the financing emerged, people focused less on the amount, valuation or investors and more on one question: why has DeepSeek suddenly started raising money?

For now, talent may be one reason.

Over the past year, the fiercest competition in the large-model industry has gradually shifted from model capabilities to talent.

OpenAI, Meta and China’s leading model companies are all continuing to compete for top researchers and engineering talent. For DeepSeek, fundraising will undoubtedly give it more substantial incentive tools to help retain its core team.

But retaining talent alone may not be enough to explain DeepSeek’s recent moves.

Recently, Huxiu noted that DeepSeek is hiring for a “senior data center delivery manager” role. According to the job description, the position covers end-to-end management of data center projects from initiation, construction and delivery to operations, while also requiring involvement in building IDC automated operations platforms, resource management and operations standards systems.

More notably, DeepSeek specifically cites experience in delivery and operations tied to GPU computing power, as well as experience planning and building large-scale clusters, in the role requirements.

For an ordinary internet company, this might simply be an infrastructure role. But for a large-model company, it points to a different capability system.

Over the past two years, the industry has talked most about model capability. R1, V3 and V4 each drew outside attention whenever they were released. But as model sizes keep expanding and token consumption driven by agents continues to rise, model companies no longer face only the question of how to train a model.

More and more problems are emerging outside the model itself. How should increasingly large GPU clusters be managed? How can computing power utilization be improved? How can training and inference tasks be kept stable? And how can companies support the infrastructure scale required to train the next generation of models?

These are exactly the problems data center teams are meant to solve. At the same time, DeepSeek is also continuing to expand its harness team.

For many model companies, the harness is closer to a training and inference infrastructure platform. It does not directly take part in model research and development, but it helps training, inference and resource scheduling run more efficiently.

Seen from this angle, DeepSeek’s current focus is no longer limited to model capability. It also includes infrastructure capability, because that determines the upper bound a model can reach.

That may be one of the most important changes to watch after DeepSeek’s financing.

One Flow Goes to Users, the Other to GPUs

2026 is a critical year for Kimi. That much is clear from the fact that it raised money three times in six months and saw its valuation rise fivefold.

At the start of the year, OpenClaw delivered striking revenue growth. For the first time, Moonshot AI saw a real path to commercialization.

At the same time, changes in several models Kimi released this year show that model technology is increasingly being designed to serve products.

Whether in agents, coding or overseas markets, Moonshot AI is trying to prove that model capabilities can ultimately translate into revenue and growth.

For Kimi, that means it must continue pushing model capabilities higher while accelerating commercialization. DeepSeek’s focus is different.

Although both companies are raising money and expanding their teams, the problems they need to solve are completely different.

For Kimi, the market cares about technology, but it cares even more about the growth brought by technology and products. Can agents create new paid use cases? Can overseas markets replicate the success seen at home? These questions all ultimately lead back to commercialization.

DeepSeek faces a different pressure. Since its founding, DeepSeek’s greatest competitiveness has always come from model capability. Whether it was the low-cost training of the V2 era or the boom in reasoning models triggered by R1, its core advantage has been built on technological leadership.

For DeepSeek, then, the most important question is not how to find more users, but how to keep staying ahead.