What developers really think about Qwen3.8-Max: “An API business model wearing an open source jacket”
The New Stack Adrian Bridgwater ● Covered by 6 sources
Alibaba dropped Qwen3.8-Max, a 2.4-trillion-parameter model it calls open source, but the weights aren't out yet. Developers say the benchmarks are self-graded and the real story is the harness built to babysit it.
Alibaba spent this week's launch spotlight on Qwen3.8-Max, its biggest model yet at 2.4 trillion parameters, built on the Qwen 3.5 architecture with a million-token context window and a sparse mixture-of-experts setup meant to keep inference costs down. The company also promised, for the first time, to open-source the weights of a Qwen-Max-class model. Promised, not delivered — the weights land next week, not today.
That gap between announcement and delivery is exactly what caught the attention of developers The New Stack talked to. Jeff Brokaw, an independent AI consultant who's been shipping software since 2016, called it "the API business model wearing an open source jacket for the launch photo," noting Moonshot AI ran the same play with Kimi K3 months earlier. He's also skeptical of Alibaba's own numbers: Qwen3.8-Max ranks fifth in text and second in vision, but those scores came from Alibaba's own same-day evaluation against a comparison set Alibaba picked itself. Brokaw's read: that's ambition dressed up as a leaderboard, not an independently verified result.
Other engineers zeroed in on what the model actually did rather than what it claims. In internal testing, Qwen3.8-Max ran a real software engineering project autonomously for sixteen days, producing 265 commits and 127 pull requests, gated the entire time by an issue-tracking system, a dispatcher, a monitor, and continuous integration checks. Anwar Khan, a senior engineer at John Deere, points out that Alibaba itself didn't trust the model unsupervised — it trusted the cage of tests and checks surrounding it. He argues that's the part worth replicating, not the raw autonomy claim.
The million-token context window drew similar pushback. Khan and others note that stuffing a huge window with data doesn't guarantee the model finds the right passage — it just increases the odds of a confidently wrong answer buried in noise. What the giant window really buys, several engineers agreed, is faster prototyping: skip building retrieval infrastructure and still get a working demo. That convenience quietly breaks down once a real corpus outgrows the demo.
Even Alibaba's own onboarding material undercuts the idea that Qwen3.8-Max is about to unseat Claude or GPT. Khan points out that the quickstart guide tells developers to keep using Claude Code or Codex and simply repoint the API endpoint to QwenCloud. As Khan put it, even the challenger assumes your existing tools stay exactly where they are.
My take
Calling something open source when the weights ship a week later, on Alibaba's own schedule, isn't a technicality — it's marketing. Same goes for grading your own benchmarks the day you launch and calling it a ranking. The sixteen-day autonomous coding run is the actually interesting bit, and predictably it's the part getting the least attention, because it doesn't fit neatly into a leaderboard screenshot. Enterprises adopting this stuff should copy the guardrails Alibaba built around its own model before they copy the hype about its parameter count.
Read more about this at: The New Stack
