On the last weekend of August, two Chinese labs told you exactly opposite things about the future of open weights, and both were right.
Alibaba spent the month teaching the market a phrase. Its Qwen3.8-Max weights — roughly 2.4 trillion parameters, the biggest and most capable model it had ever released openly — shipped not under Apache 2.0 but under a custom license with a hard commercial gate: any Model-as-a-Service or AI Work Assistant business clearing $50 million in trailing revenue has to negotiate a separate agreement, per Forkast. Moonshot had set the template weeks earlier with Kimi K3, reportedly free below $20 million of trailing-twelve-month group revenue and, above that line, a paid deal sources put at up to 30% of revenue.
The market compressed all of it into four words: open-source AI is no longer free.
Then, reportedly on August 29, Tencent shipped Hy4 preview — 770 billion total parameters, 49 billion active, a million-token context window, 1.56TB on Hugging Face and, per Simon Willison, materially bigger than the 295B Hy3 it replaced. It arrived with no revenue-share clause at all. Free for two weeks on WorkBuddy and CodeBuddy. Free Hy3 access extended to September 30 on the same platforms.
A bigger, more capable open model, released the same weekend the "no longer free" narrative hardened into consensus, asking for nothing. That is not a footnote to the revenue-share story. It is the refutation of the idea that there is one story.
Last week in this space we tracked the DeepSeek price floor — the argument that open weights had put a hard ceiling on what anyone could charge per token. This week the frame shifts from capability and price to the license text itself, because the license text is where a lab discloses its actual business model. And read side by side, Alibaba, Moonshot, and Tencent have not agreed on a toll gate. They have split over where to put it.
Start with what the revenue-share camp is actually doing. The reflexive reading — labs taxing their own giveaways — gets the mechanism right and the target wrong.
The meter isn't installed on the weights. Run Qwen on your own hardware and it's free. Build an app on somebody's API and you pay ordinary token prices. The clause only fires when you become a Model-as-a-Service provider — when you resell the model's output at scale, which is precisely Alibaba's own API business.
Alibaba reportedly prices the Qwen3.8-Max API at $2 input and $6 per million output, holding parity with GPT-5.6. The $50 million gate exists to make sure nobody undercuts that API using Alibaba's own weights. Forkast called it correctly: not a gift, a platform play. The license creates a safe harbor for startups and forces exactly the companies large enough to threaten the API business to the negotiating table.
So the revenue-share license is not "open weights becoming closed." It is a lab metering the model itself because the model is the product it sells. The toll gate sits on the thing you downloaded, and it opens for you until you get big enough to compete with the seller.
The toll gate sits on the thing you downloaded, and it opens for you until you get big enough to compete with the seller.
Tencent's Hy4 puts the gate somewhere else entirely — and the giveaway is the point.
Reuters reported that Tencent plans to integrate Hy4 alongside CodeBuddy and WorkBuddy, the developer and productivity products it already sells to enterprises. The two-week free trial isn't generosity; it's a funnel. Try the frontier model free inside Tencent's applications, then keep paying Tencent for the applications, the cloud, and the support around it.
Where you can pay for the raw model, Tencent prices it low: $0.834 per million input tokens and $2.501 output on Tencent Cloud TokenHub, per Tencent — well under Alibaba's Max API. That's the tell. If your money comes from the surrounding stack, you want the model as cheap and as widely deployed as possible, because every deployment is a lead into CodeBuddy, WorkBuddy, and Tencent Cloud.
Both labs are extracting revenue at scale. But one has bet its income statement on the weights being valuable enough to meter, and the other has bet on the weights being a commodity worth giving away to sell everything adjacent to them.
These are rival theories, not two flavors of one trend. And the license text tells you which theory a lab has wagered its business on before any benchmark does. A lab that thinks the model is the moat writes a revenue-share clause. A lab that thinks the moat is the product surface and the cloud writes Apache-style terms and a free trial.
Which brings us to the buyer, and the uncomfortable part.
RunPod has been making the cleanest version of the case for owning your weights: renting a model means your millionth request costs what your first one did, while owning the weights makes cost per request something your engineers can lower — through quantization, distillation, batching policies you control. The rate can also change out from under you. When DeepSeek raised API prices on August 16 by anywhere from 50% to more than 1,100% depending on the model, token type, and hour, anyone already running those weights on their own hardware paid exactly what they'd paid the day before, per RunPod.
That is a real advantage — and last week's DeepSeek repricing shows even the royalty-free camp can jack up the rental at will, which only sharpens the case for keeping the weights yourself.
But look at what the revenue-share license does to that escape hatch.
Owning your weights beats the API only when the lab lets you keep the weights cheaply. Under a revenue-share license, the moment your unit economics start working — the moment you're big enough that self-hosting genuinely undercuts the API — is exactly the moment the license kicks in and drags you back to the negotiating table. Moonshot's $20 million line. Alibaba's $50 million line.
The revenue-share tier is engineered to close the RunPod escape hatch at precisely the point it becomes worth using. "Own your weights" stops being a permanent strategy and becomes one with an expiration date — and the date is set by someone else's income statement, not yours.
That's the real division underneath the "no longer free" panic. It's not open versus closed. It's whether the lab you built on wants the weights to be a royalty-bearing asset or a lead-generation loss leader — and only one of those answers leaves your cost curve in your own hands past a certain scale.
Which is why Hy4 shipping with no revenue-share clause is more than a competitive jab at Alibaba. It's Tencent telling large-scale builders that with Hy4, the RunPod math holds all the way up — no clause waiting to fire when you grow into it — because Tencent intends to make its money somewhere you're going to spend it anyway.
For a buyer choosing a foundation this quarter, the benchmark deltas are close enough to argue about. The license terms are not. One camp has priced your future success into the contract. The other is betting it can sell you the rest of the stack once you're hooked.
Read the license before you read the leaderboard. It's the only document that tells you what happens to your margins the day your product actually works.