StepFun plays hardball after soft approach fails

"Waiting for Qi Yin to Make His Move"

Of the model companies still in the game, StepFun is probably the weirdest. Its model capabilities are just okay — broadly agreed to be second-tier.

It does occasionally pull off a surprise. Step 3.7 Flash, which our colleague tested on Xianyu, had insane price-performance.

But 3.7 is an old model now. Lately Hy3, MiniMax M3, LongCat 2.0 and other domestic models have all slashed prices dramatically — 50% off at minimum — so StepFun can't even get a seat at that table anymore.

But honestly, nobody cares about that now.

Six months ago, when we wrote "The Four AI Tiger Cubs Enter the Hardcore Competition", our core thesis was that for large model companies today, both 2B and 2C are brutally hard.

For 2B, customers have Alibaba Cloud and Volcano Engine to choose from. The only major variable since then is that Zhilin Yang and Tang (MiniMax's founder) have used SOTA model capabilities to prove, at least for now, that model companies can do 2B.

For 2C, Doubao owns the user entry point. And judging by Doubao's attempt to charge, Chinese consumers paying for C-end products is... a bit of a joke. The C-end battlefield has shifted to on-device agents — feels like 100 local agent startups are fighting in the mud.

So StepFun's story for the past year has been hardware closed loops, binding to scenarios, software-hardware integration, then going 2C together.

When Qi Yin first emerged six months ago, he was already talking about two-wheelers, four-wheelers, bipedal robots, phones, cars, robots — full coverage.

Now he's rolled out STEPX, Step AOS (agent-native operating system), Amoo (agent), STEPX Neo (phone), StepFun AI desktop companion, and Super EVA car agent.

I was laughing my ass off at the phone launch event. Saw the Amoo logo and thought I was seeing things — did Anthropic break into China?

This thing is literally the Anthropic logo with a bangs fringe added on.

I asked around — the trademark is already registered. So you're bullying Dario because he can't file an IP lawsuit from across the Pacific? This is peak Xiaomi energy.

What impressed me most was when Qi Yin unveiled the large-model-native AI terminal brand STEPX later in the event. For a second I thought Elon Musk had come to China.

I'm speechless. Nothing to say except respect. Felt like he was about to shoot a Musk multiverse movie any second.

Let me break down StepFun's grand chessboard:

StepFun itself has full-modality models — agent models, multimodal models, edge models, plus StepFun voice models (not sure if this is mainly for getting into cars).

Closely tied to it, Qianli and Geely Auto bring vehicle platforms, engineering, manufacturing and distribution channels. They've also built a car-ready large model to compete with Doubao getting into cars.

They've also invested in a humanoid robot startup called Yuanling Robotics.

Now they've added the smart agent phone segment, plus a native OS for agents.

Normally a model company could never make phones solo — lacking hardware experience, the supply chain would destroy them. Phone makers doing AI phones, meanwhile, get stuck on their own models.

But for Qi Yin, who's already integrated models, system, and terminal as one, this is obviously no problem. He doesn't talk about ecosystem as much as other companies do daily, but he's integrated faster than anyone — blink and he's cranked out a pile of hardware.

I'm making a direct wish to Qi Yin right now: one AI PC, please. Hoping to receive it in six months.

At WAIC, three companies launched AI phones: StepFun, Doubao, and HONOR, all claiming the world's first AI agent phone.

HONOR's Robot Phone features a gimbal — like they stitched an Insta360 inside;

Doubao phone is making another attempt. Wonder if this time they'll integrate Douyin, local services, Lark and their own services;

StepFun phone's focus is Step AOS. Its capability layers are memory, decision-making, and action.

Memory solves cross-app, cross-device information processing. Decision-making combines on-device fast response with cloud complex reasoning. Action, of course, is the hardest part.

StepFun's current plan is to get apps to open APIs and authorize access, letting Step AOS operate them directly. This would effectively solve the GUI problem, but the puzzle is: why would others agree?

The classic negative case here is LightSail headphones — claimed you could order by talking to them, but in reality could only photograph products and dump them into JD.com's cart, no checkout permission.

Hoping Qi Yin, who can best Gaoding ecosystems, makes his move fast.

Only while writing this did I realize how ridiculously good Qi Yin's story is.

Phones are closer to daily life than desktops — travel, payments, photos, schedules, communications, everything. For StepFun, the phone carries data feedback loops, users' work flows and life flows, helping the model understand real tasks.

Think longer-term: the phone could pair with a detachable small hardware combo, just slap a Looki on top, dedicated to recording video and information capture. Hell, this phone could eventually become a portable compute center.

Can only say Qi Yin has never lost a game of chess.

StepFun phone recently launched a wishing campaign — meaning you can wish for future features.

Here I go:

Smartphones must be smart. AI phones must understand AI. After this StepFun phone launches, it better deploy a small local model. Capability bar isn't high — just stronger than ERNIE Bot.

If it can't even do that, send it to mine Bitcoin;

Whatever demands people made of Doubao, you gotta match at minimum. Auto-calling to reserve Sushi Lang, Kaojiang, and Chef Fei's stir-fried pork with chili. Better yet, book us a venue for the next cybercafe hackathon.

If you're worse than Doubao that's just pure comedy.

This phone also gotta understand sign language. Take it to Northeast China, do the flower-hand dance, it better recognize which MC star you're repping — that's how you save face.

Finally, giving away a Claude figurine in the comments, ah no, I mean a StepFun Amoo figurine.

(Cover image generated by ChatGPT, article 100% human-written)

Subscribe to our Substack: funeralai.substack.com