What developers really think about Qwen3.8-Max: “An API business model wearing an open source jacket”
The New Stack Adrian Bridgwater ● Covered by 10 sources
Alibaba launched Qwen3.8-Max, a 2.4-trillion-parameter model, but the open weights aren't out yet—just promised for next week. Developers say the benchmarks are self-graded and the real test is the harness, not the parameter count.
Based on reporting by The New Stack, Adrian Bridgwater — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Alibaba spent this week talking up Qwen3.8-Max, the biggest model in its Qwen lineup so far. It's multimodal, built on the Qwen 3.5 architecture, and packs 2.4 trillion parameters alongside a context window that stretches to a million tokens. Under the hood it uses a sparse mixture-of-experts setup and hybrid attention, both aimed at making that scale actually workable rather than just impressive on paper. Alibaba also said this marks the first time it will open-source the weights for a Qwen-Max-class model — except it hasn't done that yet. The weights are due next week.
That gap between announcement and delivery is exactly what caught the attention of developers The New Stack spoke to. Jeff Brokaw, an independent AI consultant who's been building software since 2016, pointed out that Moonshot AI ran the identical move with Kimi K3, which itself launched with 2.8 trillion parameters. Promise the open weights, generate the launch buzz, ship later. Brokaw called it an API business model wearing an open source jacket for the photo op. He was equally skeptical of the rankings Alibaba published — fifth in the text arena, second in vision — noting they came from Alibaba's own same-day evaluation against a comparison set Alibaba chose itself. His read: treat it as ambition, not a settled result, until someone outside the company runs the same tests.
Other engineers zeroed in on different pressure points. Phil Whittaker at Umbraco argued that as models get commoditized, benchmarks matter less than speed, token efficiency, and — above all — the quality of the harness wrapped around the model. He said the rankings wouldn't push him to switch away from established players like Claude and OpenAI, since integration and harness quality still decide who wins in practice. Anwar Khan at John Deere zoomed in on a detail buried in Alibaba's own writeup: an internal test where Qwen3.8-Max ran a real software engineering project autonomously for 16 days, producing 265 commits and 127 pull requests, all gated by an issue tracker, a dispatcher, a monitor, a watchdog, and continuous integration checks. Khan's point wasn't that the model earned trust — it's that Alibaba built a cage around it and trusted the cage, not the model. He called that the correct instinct and the part worth copying.
The million-token context window drew its own scrutiny. Khan warned that stuffing a huge context window with material doesn't improve retrieval — it just makes the right passage compete with a lot of similar-looking noise, and the failure mode is a confident wrong answer rather than an obvious refusal. What the big window does buy, he said, is prototype speed: skip building retrieval infrastructure and still get a working demo. Heath Squier of Evkii framed the bigger picture — Qwen3.8-Max paired with open weights and Alibaba's own Qoder tooling is an attempt to hand enterprises a full alternative stack for running agents, not just a model. But he cautioned that a 16-day autonomous run changes the scale of risk, since small mistakes can compound across thousands of actions before anyone notices.
Maybe the most telling detail came from Alibaba's own quickstart guide. It tells developers to keep using Claude Code or Codex and simply point the base URL at QwenCloud. As Khan put it, even the challenger assumes your existing tooling stays put.
My take — AI-written commentary, not fact-checked reporting
Nobody serious is going to rip out their Claude or OpenAI setup because a lab published its own same-day leaderboard placement — that's marketing, not evidence, and every developer quoted here saw through it instantly. The more interesting story is that Alibaba, like Moonshot before it, is learning that dangling "open weights coming soon" gets you launch-day headlines without the inconvenience of actually shipping anything. The real signal buried in this release is the 16-day autonomous coding run wrapped in a heavy monitoring cage — that's the part enterprises should study, because it proves even the vendor doesn't trust the model on its own.}
Read more about this at: The New Stack