China's open-model machine is not slowing down. It is speeding up.

In a single week, three of China's biggest labs put three frontier models on the board. Alibaba previewed Qwen 3.8 and claimed the number-two spot in the world. DeepSeek is shipping the stable release of V4 this Friday. And Zhipu's GLM 5.5 is already leaking for August. All of it while Google's Gemini 3.5 Pro still cannot make it out the door.

The pattern matters more than any single model: frontier-adjacent quality, open or near-open weights, and prices a fraction of what US labs charge, arriving faster than anyone can benchmark. Here is what each one actually is, and what is real versus hype.

Alibaba fired the loudest shot. On Sunday its Qwen team previewed Qwen3.8-Max, a 2.4-trillion-parameter multimodal model it says is one of the most capable available today, second only to Anthropic's Fable 5. That would put a Chinese model at number two in the world, ahead of everything from OpenAI and Google. Investors liked it: Alibaba shares jumped as much as 5 percent. It is available in preview now through Alibaba's Token Plan, Qoder, and QoderWork platforms at roughly a tenth of standard pricing.

Two things keep the hype honest. First, the ranking is Alibaba's own internal claim, with no independent benchmarks, no active-parameter count, and no model card published yet, so nobody outside the company can verify the second-place boast. Second, and more interesting, is the open-weight promise: Alibaba says Qwen 3.8 will go open-weight soon, which would reverse its recent habit of keeping its strongest Max models closed and API-only. If it follows through, the most capable Qwen ever becomes a free download, which is a much bigger deal than any leaderboard position.

The most influential name in Chinese open models is up next. DeepSeek V4, which has been rolling out in preview through mid-July, moves to a stable release on July 24, and its older API aliases retire the same day, so anyone on the old model names has a hard migration deadline. V4 also brought a first for a frontier API: peak and off-peak surge pricing, charging more during Beijing business hours and less overnight, a pricing experiment no US lab has tried.

DeepSeek is the lab that kicked off this entire cheap-and-open wave, and V4 keeps the formula: frontier-adjacent quality at a fraction of Western prices, with V4-Pro previously among the largest open models on the market before Kimi K3 took that crown. For builders, Friday is the date that matters, both for the stable weights and for the aliases that stop working.

🔮 GLM 5.5 Is Lining Up for August

The third name in the pileup has not even shipped, and it is already generating heat. Zhipu, the Beijing lab that markets itself as Z.ai, is reportedly lining up GLM 5.5 for August. Nothing is official: no date, no specs, and no price from Zhipu itself, only a two-month release cadence (GLM-5 in February, 5.1 in April, 5.2 in June) that points cleanly at an August drop.

What makes the next GLM worth watching is what the current one already did. GLM-5.2, released open-weight under an MIT license in June, was independently ranked the most capable open model on the market at launch. Zhipu has publicly committed to shipping a Fable-class open model by the end of the year, and GLM 5.5 is the next step toward it, at roughly a sixth of frontier API cost. If it lands in August as expected, it extends the exact pattern Qwen and DeepSeek just set this week.

For a year the comfortable story was that the US owned the frontier and China owned the discount rack. That line is getting harder to say with a straight face. Three Chinese labs just claimed frontier-adjacent performance in a single week, two of them promising open weights you can self-host, at roughly a tenth of Western API cost. The US still holds the very top, Fable 5 and GPT-5.6 Sol remain ahead, and every one of these Chinese claims is unverified until independent benchmarks land. But the direction is unmistakable, and the cadence is the weapon: while US labs ship one flagship a quarter and Google slips its deadline yet again, China's labs are shipping monthly. For builders, the move is to actually test these on your own workloads now, because the assumption that the good models only come from US APIs is aging fast.

The stories moving fast beyond today's headlines:

China's open-model machine had its loudest week yet. Alibaba previewed Qwen3.8-Max, a 2.4-trillion-parameter multimodal model it claims is second only to Anthropic's Fable 5, sending its stock up as much as 5 percent, though that ranking is Alibaba's own internal number with no independent benchmarks or model card yet, and the bigger story is its promise to release open weights soon after keeping recent flagships closed. DeepSeek moves its V4 model to a stable release on July 24, retiring older API aliases the same day and cementing its first-of-its-kind peak and off-peak surge pricing. And Zhipu's GLM 5.5 is already leaking for an August drop, building on GLM-5.2, which was independently ranked the most capable open model at its June launch. The unifying thread is cadence and cost: three Chinese labs claiming frontier-adjacent performance in a single week, two promising open weights you can self-host, at roughly a tenth of Western API prices, all while Google's Gemini 3.5 Pro slips its deadline a third time. The US still holds the very top with Fable 5 and GPT-5.6 Sol, and every Chinese claim stays unverified until independent numbers land, so the honest read is not that China has won, it is that China is closing the gap faster than the leaderboard's top line suggests. For builders, the move is simple: test these on your own workloads now, because the assumption that the best models only come from US APIs is aging by the week.

Stay building. 🤖