Four Chinese Models in Eight Weeks, and Not One of Them is Doing the Same Thing

Between 13 June and 3 August, four Chinese AI labs each released a flagship model. Zhipu released GLM-5.2, Moonshot released Kimi K3, DeepSeek published a rebuilt version of V4-Flash, and Alibaba launched Qwen3.8-Max.

Almost all of the coverage has treated this as a race with a scoreboard and asked which one won. A ranking is a poor guide here, because the four differ in ways a score cannot show, and those differences are what decide whether a company can realistically put a model into production.

Three differences do most of that work. The first is whether you can run the model on your own machines or only rent it over the internet. The second is what it costs, where the gap between these four is more than fiftyfold. The third is whether anyone outside the company that built it has checked the numbers.

Here is where each one sits.

Kimi K3, from Moonshot AI, is the most capable and the least practical. Launched on 16 July, it has 2.8 trillion parameters, the internal values that determine how a model behaves and the industry's rough shorthand for scale. Size alone proves little, so the more useful figure comes from Artificial Analysis, an independent outfit that runs the same tests across every model and publishes one combined score. It puts Kimi K3 at 57, the highest any openly published model has recorded. The model also ranks first on one of the main public leaderboards for front end coding.

Moonshot published the model files on 27 July. They run to roughly 1.56 terabytes across 96 separate files, which puts self hosting out of reach for anyone without serious infrastructure. The licence is also its own document rather than one of the standard permissive ones, and it carries two conditions. Companies reselling access to the model need a separate agreement with Moonshot once group revenue passes $20 million over any twelve months. Products with more than 100 million monthly users, or more than $20 million in monthly revenue, have to display the name Kimi K3 in the interface. Neither condition affects a small team, but both need a lawyer's eye at enterprise scale.

GLM-5.2, from Zhipu AI, is the one most companies can actually deploy. Released on 13 June at roughly 744 billion parameters under the MIT licence, which imposes almost no conditions at all, it scores 51 on the same independent index. That is meaningfully below Kimi K3 and roughly level with several current American models.

Its adoption tells the more interesting story. Vercel, a hosting company that routes AI requests for over 200,000 development teams and publishes monthly data on what they use, recorded GLM-5.2 as the fastest uptake of any model it tracked in 2026. Daily volume grew around fiftyfold in its first two weeks, and it reached eleventh place on the platform by usage within a fortnight of release. Teams did not wait for consensus before trying it.

DeepSeek V4-Flash is the price story. AI companies bill by the token, a fragment of a word, sold in batches of a million. DeepSeek charges $0.14 for a million tokens of input and $0.28 for output, a fraction of everything else here. Kimi K3 charges $3 and $15 for the same volumes, which makes its output more than fifty times more expensive.

What makes V4-Flash worth attention is what DeepSeek did on 31 July. It republished the model with no change to its size, its price or its licence, having simply retrained the final stage. Its score on the independent index went from 40 to 50. That is a substantial capability gain delivered without a bigger model or a higher price, and it is the clearest evidence this year that scale is no longer the only lever worth pulling. DeepSeek is now the third largest source of AI traffic on Vercel's platform at 22.6 percent, behind only Anthropic and Google.

Qwen3.8-Max, from Alibaba, is the one to check twice. Released on 3 August at 2.4 trillion parameters and priced at $2 and $6 per million tokens, it arrived with a claim from Alibaba that it is second only to Anthropic's Fable 5. That claim is now testable, and the answer depends entirely on which test you look at. On Arena's board for image and video work, Alibaba is right and the model ranks second. On the board for text, it ranks fifth.

Both numbers are real and they point in different directions, which is a reminder that a model is not one thing with one ranking. The rest of Alibaba's published figures, including a 16 day autonomous coding run it describes in detail, are still its own and have not been independently reproduced.

Alibaba also said it would publish the model files during the week of 10 August, which would be the first time it has done so for a model at this tier. As of the most recent reporting, nothing had appeared and no licence had been named. That gap is worth watching. Every previous top tier Qwen model has stayed closed, meaning the files were never released and the only way to use one was to pay Alibaba for access over the internet.

Four models, four different definitions of "open"

"Open" is being used for all four of these, and it means something different each time.

DeepSeek and Zhipu published their models under standard permissive licences, so a company can download them, run them on its own hardware and build a commercial product without asking anyone. Moonshot published its files too, but attached conditions that apply only above a certain size, so a small team is unaffected while a large one needs legal sign off before committing. Alibaba has published neither the files nor a licence, only a stated intention to release both.

The practical difference matters more than the ranking. Running a model on your own hardware means your data stays inside your company, and the vendor cannot raise your price or withdraw the model. Renting one over the internet means the vendor controls all three, whatever the model scores.

What this is worth to a business

If you are choosing between them, the honest answer is that the choice depends on the job. GLM-5.2 is the safe default for a team that wants to self host, because it is capable enough and its licence creates no future problems. DeepSeek V4-Flash is the right choice for high volume, lower stakes work where the cost difference compounds. Kimi K3 is the most capable, and only realistic through its interface unless you have a data centre. Qwen3.8-Max should not go into a plan until its files and its licence exist.

If you are watching this from a marketing or brand position rather than an engineering one, the relevant point is that these are the models increasingly answering questions about your products inside other companies' software. Which one is underneath any given assistant is rarely disclosed, and as this list shows, they do not behave the same way.