Alibaba announced on 3 August 2026 Qwen 3.8-Max, a 2.4-trillion-parameter model touted as bringing improvements in coding, office work and research. The weights are due to be released the following week. A week after Kimi K3's actually went live, the race for openness is heating up among Chinese labs.
In late July, we documented the release of the weights for Kimi K3, then the largest open model ever distributed. A week later, a competitor announces a model of comparable size with the same promise. Let's look at what this means.
What is being announced
Qwen 3.8-Max boasts 2.4 trillion parameters, placing it in the same category as the largest current open models. Alibaba highlights improvements on three fronts: programming, collaborative office work, and document search.
This positioning is telling. It's not about claiming the best overall performance, but targeting the uses where businesses actually spend money. Code and administrative work account for most of the workloads billed today, as we saw with the ATLAS study.
The promise to open the weights the following week deserves the same caution we applied to Kimi K3: an announced date is not a downloadable file. Moonshot kept its commitment on time, setting a favourable precedent, but each promise is judged when it comes due.
As with any launch announcement, the performance claims come from the lab itself and have not been independently verified at the time of writing. The parameter count, moreover, says nothing on its own: you'd need to know the share actually activated per request, as we explained regarding mixture-of-experts architectures. A 2.4-trillion-parameter model where 40 billion work at any given moment does not have the same cost profile as an equivalent dense model.
What this accumulation reveals
In three weeks, we've covered Kimi K3 and its 2.8 trillion parameters, a DeepSeek update at 14 cents per million tokens, and now Qwen 3.8-Max. Three Chinese labs, three major announcements, two openness promises kept or in progress.
This pace produces a mechanical effect we now observe every week: it pushes prices down everywhere. The 80% price cut announced by OpenAI can't be explained solely by technical gains. It fits into a market where a competitor offers similar capabilities for a fraction of the price, or even free to download.
For users, this is unambiguously good news. For the sector's balance, the question remains open: an industry where the core product becomes free must find its profitability elsewhere, and no one yet knows where.
What to take away
The important point is not the parameter count, which will become mundane in six months. It's that openness has become a systematic competitive argument rather than a militant exception.
A year ago, publishing the weights of a state-of-the-art model was a political gesture. Today, it's an expected box on a product announcement, to the point that a lab that didn't do it would have to explain itself. This shift, which we detailed in our article on open models, is arguably the most structural development of 2026. The question is how long an industry can sustain this pace while giving away its product.