These days, it seems like a gathering of Chinese large-scale models is underway. First, K3 made headlines, and then Qwen3.8 was unveiled yesterday, both boasting models with over 2T parameters. It's rumored that DeepSeek V4 will also be launched today. Initially, I planned to discuss it after V4's release, but the anticipation was just too high to wait.

However, even with Qwen's current moves, there are already numerous insights worth delving into.

In terms of timing, Kimi K3 has just been launched and has performed exceptionally well. DeepSeek V4 is on the verge of release, and while its capabilities remain a mystery, based on Liang Sheng's usual standards, it's certainly something to look forward to.

Caught between these two, Qwen3.8 finds itself in a somewhat precarious position. If its model capabilities can't surpass K3 and its cost-effectiveness can't outdo V4, external promotion will be an uphill battle.

Another noteworthy aspect is that Alibaba explicitly stated its intention to open-source Qwen3.8 at the time of its release. (The author of the WeChat official account post even described the Qwen large model as 'poised for open-sourcing.')

One factor contributing to Lin Junyang's departure was the group's adjusted stance on open-sourcing, beginning to weigh the costs and benefits more heavily.

Alibaba had shown signs of this shift earlier, transitioning from 'actively embracing open-sourcing' to 'conditionally embracing it' and then to 'selective open-sourcing.' The open-sourcing commitment for Qwen3.8 Max indicates that Alibaba is reverting to its main strategy of the past three years.

The Art of Translation and Wording

Interestingly, there's a subtle difference in the introductions published by Qwen's official Chinese and English accounts. The English version boldly states 'second only to Fable 5.'

However, the Chinese version surprisingly adds a caveat, saying, 'it may be the most powerful model apart from Fable 5.' This raises questions. Does it imply that the officials are not entirely confident in Qwen3.8's capabilities, or are they wary of domestic overhyping that could backfire?

Qwen still seems to lack a bit of confidence.

Among domestic companies developing large models, Alibaba is relatively prone to overhyping. This is related to its management style; Alibaba is also the most optimistic during earnings calls.

The most typical example is the previous 'Happy Horse,' which dominated benchmarks upon release but fell short of expectations when users calmly evaluated it later. The initial benchmarking was so strong that it somewhat tarnished the perception.

Of course, the outcome was positive, at least recognized by Alibaba's top management, as Zhang Di, the person in charge, now oversees the group's multimodal visual business line, with Wanxiang, Happy Horse, and Happy Oyster all merged into the Future Life Lab.

Perhaps having learned from past mistakes, Alibaba dares not make absolute claims in promoting its foundational models this time.

Currently, the addition of the qualifier 'possibly' seems artistically and reasonably done.

Because most initial community evaluations conclude that Qwen3.8's current performance falls short of the preemptively released K3. So, if this qualifier hadn't been added in the Chinese promotion, the online ridicule would likely be unimaginable now.

However, officials also emphasize that Qwen3.8 is continuously evolving daily, meaning post-training is not yet complete.

As we all know, the potential of post-training is still largely unknown. Referring to Tencent's HY3 official version's improvement over the preview version, it's not impossible for Qwen3.8's official version to follow a similar trajectory and achieve a comeback.

Moreover, even if Qwen3.8 doesn't immediately outperform K3 in user experience, Alibaba still stands to benefit.

Alibaba holds over 30% of Yuezhiyan's equity, and the computational power for training K3 mainly comes from Alibaba Cloud. So, in any case, Alibaba comes out ahead: if you use K3, Alibaba Cloud earns rental fees, and its equity appreciates; if you use Qwen, Alibaba secures its open-source and cloud ecosystem positions.

This balanced capital and computational power layout is something standalone startups can only envy.

The advantage in computational power and infrastructure is directly reflected in the confidence of commercial pricing.

Kimi just announced yesterday that it had to suspend member access due to computational constraints. This exposes the harshest side of large model competition. The model itself is a trade-off between cost and intelligence. Unicorns lacking foundational cloud infrastructure quickly find themselves in a bind amid sudden traffic surges.

Meanwhile, Alibaba's Qwen3.8, though limited in capability, is fast and cheap. At a time when Kimi is struggling, Qwen3.8 is offering limited-time 90% discounts and even nighttime double discounts similar to traditional utility peak-valley pricing.

But the question remains: why did Alibaba choose to release Qwen3.8 at this node, knowing that much training work remains and post-training is still evolving daily?

The already released K3 and the upcoming official release of DeepSeek V4 are certainly key influencing factors. If Qwen3.8 delays further, the uncertainties it faces become too great.

Internally, Alibaba likely has a relatively clear assessment of its model's true capabilities. After all, management is not naive; in an industry where benchmark-brushing is common, they are unlikely to be completely misled by team-presented data, as happened with Zuckerberg's release of LLaMA4.

After K3's release, Alibaba likely tested it and found that its Qwen3.8 did not have an absolute advantage in capability.

On the other hand, DeepSeek is known as a price cutter, with Liang Sheng's reputation for extreme engineering optimization and extremely low inference costs hanging over everyone's heads.

This creates a special time window.

If Alibaba sticks to its guns and waits for Qwen3.8 to be fully trained before releasing it after DeepSeek V4, the situation will become passive.

Even if V4's official version has average model capabilities, as long as DeepSeek maintains its cost-cutting approach and slashes prices, Alibaba will be left with neither advantage: in terms of absolute model capability, you can't surpass K3, which has already occupied the high ground in reputation; in terms of cost-effectiveness, you can't beat DeepSeek V4.

Even if Qwen3.8 is a well-rounded model, once it falls into the 'inferior in capability and price' dilemma, promotion will completely lack selling points, leading to an extremely awkward situation.

Qwen's Max Version Returns to Open Source, Lin Junyang's Departure Seems in Vain

Lin Junyang's recent departure from Alibaba sent shockwaves through the large model community. There were many rumors about his departure, one being the severe internal disagreement and strategic struggle within Alibaba over open-source and closed-source approaches.

Lin Junyang represented Qwen team's early technological idealism (possibly also considering personal achievements). Under his leadership, Qwen surged ahead with unreserved open-sourcing, earning a high reputation in the global developer community.

However, this purely charitable open-source approach, as model parameters soared to hundreds of billions and trillions, inevitably clashed with the group's financial books. Training a top large model requires astronomical computational power and funds, and group executives cannot ignore ROI.

A report by 'LatePost' mentioned that Alibaba did not want to open-source Qwen Max, but Lin Junyang also wanted to promote open-sourcing.

Open-source community reputation has its benefits, but they are intangible.

From the perspective of some group executives, open-sourcing the top flagship model without reservation not only fails to directly generate API revenue but also amounts to spending real money to create weapons and giving them away to the entire industry, even competitors, allowing them to poach Alibaba Cloud's customers.

In the period after Lin Junyang's departure, Alibaba's large model strategy underwent significant contraction and reorientation.

The open-source pace for flagship models slowed, shifting to a utilitarian approach of 'open-sourcing small and medium models while keeping large models closed-source.'

Alibaba's reluctance to open-source its largest flagship model was once seen as a signal of its gradual compromise toward closed-source commercialization and the decline of open-source idealism.

For example, the Qwen3.6 release page mentioned, 'We will also open-source smaller model versions to reaffirm our firm commitment to technological inclusivity and community-driven innovation.'

Why did Alibaba explicitly announce the open-sourcing of the 2.4T parameter Max top version with Qwen3.8's release this time? The logical shift from tightening to fully reopening is not difficult to understand; it's a correction back to common sense.

Simply put, in the large model circle now, the situation is: if you don't open-source, plenty of others will.

At this stage, the essence of open-sourcing is using the ecosystem and reputation to compensate for gaps in absolute model capability.

Imagine if your model capability truly rivals or surpasses OpenAI or Anthropic, alternating in leadership. Then, even if closed-source, everyone would line up to pay for your closed API for the best results.

But currently, open-source models still lag behind closed-source models, at least never surpassing the top closed-source models of the same period.

In the future, if independent large model companies like Zhipu and Yuezhiyan truly catch up to OpenAI or Anthropic, I believe they will likely shift to closed-source.

Because otherwise, leaving AGI aside, the sustainability of independent vendors becomes difficult to solve; you can't really treat investors as customers.

Alibaba's situation is different. I've never quite understood why Alibaba internally thought 'shifting to closed-source is more cost-effective.'

As a cloud vendor with resource advantages and a self-trained foundational model, open-sourcing brings little substantive loss to Alibaba.

Even if Alibaba open-sources the full weights, and others deploy them independently or small and medium cloud vendors provide services, their inference efficiency and cost per token are unlikely to match Alibaba Cloud's native deployment levels.

If they do, it only proves that Alibaba Cloud is too incompetent.

So, Lin Junyang's departure due to strategic disagreements was indeed regrettable, but fortunately, after a period of oscillation, Alibaba ultimately figured out the math.

Qwen3.8 Max's renewed open-source commitment and return to the past three years' main strategy indicate that Alibaba, as a cloud giant, has finally understood its core narrative:

Rather than scrimping on closed-source API profits like independent model vendors, it's better to generously use open-source models to build friendships and leverage computational efficiency to earn substantial profits in the 'utilities' of the AI era.

Joe Tsai said in Paris last month that the immense value of AI is certain, but which level these values ultimately converge at is uncertain. The four levels—underlying chips, cloud computing infrastructure, model layer, and application layer—are all possible. Alibaba's approach is full-stack, meaning it aims to be well-positioned at each level.

This analysis holds from an industry perspective, but probabilistically and in terms of revenue trends, Alibaba's greatest opportunity still lies in MaaS-based cloud computing business.