Article

Why the Studio Behind Genshin Impact Wants to Build a Top-Tier LLM

Why the Studio Behind Genshin Impact Wants to Build a Top-Tier LLM

The developer of Genshin Impact wants to be an AI company too. According to discussion on LINUX DO, miHoYo's founder told students at a Shanghai Jiao Tong University recruitment talk that the company's large language model aims to reach China's top tier within the next one to three years. The thread drew 4,500 views and 119 replies — and the most interesting part wasn't the ambition, but where the community thinks miHoYo will aim its model.

A caveat first: this is a second-hand report of a recruitment talk, discussed on a forum. Recruitment talks are part marketing, part aspiration — the timeline vagueness (one to three years is an eternity in AI) was the first thing skeptics in the thread flagged. But directionally, the move deserves a serious look, because a game studio entering the LLM race isn't as strange as it sounds.

Not code — characters

The consensus prediction from the thread: miHoYo won't chase the code-generation crown. Instead, it'll go after chat and lifestyle use cases, in the vein of DeepSeek's consumer success. The reasoning is simple — a game studio's unfair advantage isn't GPUs or papers, it's characters people already love.

Players have reportedly been impressed by AI-driven NPCs in the studio's recent games: characters that remember conversations, react with personality, and feel less like quest dispensers and more like inhabitants. For a company whose core competency is making millions of players emotionally attached to fictional people, a language model isn't a pivot. It's the missing technology for the thing they already do.

This is the part Western coverage of the AI race tends to miss. The standard framing is labs competing on reasoning benchmarks — but benchmarks are not products. The consumer AI market is ultimately decided by attachment: which assistant people open every day, which character they talk to, which product they miss when it's gone. A game studio that has spent a decade engineering attachment at planetary scale is not starting from zero. It's starting from the finish line of a different race.

The "living character" thesis

Think about what a top-tier game studio brings to the LLM race that a lab doesn't:

  • Character writing at scale. miHoYo employs armies of writers who craft distinct voices for dozens of characters. That's effectively the world's largest RLHF dataset for personality — thousands of hours of dialogue where every line had to stay in character, land emotionally, and survive contact with a fandom that notices inconsistencies.
  • Emotional attachment as distribution. People don't just use these characters; they cosplay them, write fanfiction about them, spend money on them. An AI companion from this pipeline starts with a fanbase. Most AI products have to earn attention; miHoYo's would start with it pre-installed.
  • Multimodal by necessity. Games already combine text, voice, animation, and real-time interaction. A "living" game character is the most demanding possible demo of an embodied AI — it has to look right, sound right, respond in real time, and stay in character for hundreds of hours.

There's precedent for this direction working. The global fascination with AI companions — and the willingness of users to pay for them — suggests the "characters worth caring about" market is real. Game studios have been rehearsing for that future for decades; they just called it NPC design.

Why now?

The timing makes sense for structural reasons. Inference costs have fallen far enough that a real-time conversational character is economically plausible at scale. Voice models are good enough that a character can speak in its own voice without breaking immersion. And the talent arbitrage is real: game studios employ exactly the mixture of writers, animators, and engineers that an embodied-AI product needs, under one roof.

There's also a defensive logic. If AI companions become a major consumer category, the companies that own the characters own the distribution. For miHoYo, building its own model is arguably cheaper than watching someone else's model become the definitive way to talk to its characters.

The skeptics have a point

The forum thread wasn't all enthusiasm, and the skeptical case is worth steel-manning:

  • Recruiting talks are cheap. Ambitious timelines announced to students are a genre, not a commitment. One to three years in AI is several paradigm shifts.
  • Foundation models are capital-intensive. Training a frontier-class model is a different sport from fine-tuning NPC dialogue. The thread's own members debated whether miHoYo means training from scratch or building on open weights — a distinction that changes the ambition's meaning entirely.
  • Games and AI have different clocks. A game ships when it's ready; an LLM is obsolete on a schedule. Studios that operate on multi-year production cycles have historically struggled with the relentless cadence of model releases.

None of this makes the bet stupid — it makes it a bet. The interesting question the thread kept circling is whether the "character" moat is real or sentimental. If consumers just want a smart assistant, the labs win. If they want someone, the studio with a decade of someones has a genuine edge.

The community's reaction says a lot

The forum thread itself was revealing. Alongside serious analysis, players cracked gacha jokes — "top up 648 first," referencing the game's most infamous premium currency bundle. The humor carries a real insight: miHoYo's players already have a spending relationship with AI-adjacent characters. If anyone can convert "I love this character" into "I talk to this character every day," it's them.

That conversion is the whole thesis in one sentence. The LLM race has been framed as a contest of intelligence; miHoYo's rumored bet suggests a different finish line: whoever makes AI feel alive first wins the consumer market. The labs are building better brains. The game studio is building better people.

For a lighter look at how this same community evaluates AI, see the pelican test — their unofficial benchmark involving a 3D pelican on a bicycle, which remains the friendliest entry point into how Chinese developers actually think about models.

The bigger picture: who gets to build frontier AI

The miHoYo story is part of a wider pattern: frontier AI is no longer the exclusive property of research labs. Game studios, e-commerce giants, and short-video platforms are all placing bets, and each brings a different theory of what the winning product looks like. Labs bet on intelligence; platforms bet on distribution; game studios bet on attachment. All three theories could be right at different layers of the stack — the model that wins the benchmark and the character that wins the user's evening may not come from the same company. If the next era of consumer AI is won by whoever makes it feel alive, the industry's org charts are going to look very different from today's.

Whether miHoYo's timeline holds is almost beside the point. The signal is that a studio whose business is fictional relationships now sees language models as core infrastructure, not a side experiment. That's the kind of move you make when you think the product is the relationship.

Keep reading
The Free-VPS Hunt: Inside China's Obsession With Cloud Free Tiers
Read next →