Two very different people are searching for 'qwen 3 5 397b a17b' right now. The first is scrambling to download the Qwen flagship weights before they disappear, convinced the model doesn't exist yet. The second heard rumors that 'Qwen shut down' and is frantically trying to verify whether their favorite open-source lab just died. Both are walking away confused—because the truth is more nuanced than either expected.

What Actually Exists

The Qwen3 397B A17B weights do exist, and they're sitting in the open right now under an Apache 2.0 license. No API key required. No usage limits. Just raw access to one of the most capable open-weight models available today. The confusion stems from something else entirely: this might be the last time you'll ever see flagship-tier Qwen weights released this way.

The Flagship Line Went to API

According to analysis from Russian-language AI publication Promptra, what's really happening is a strategic shift rather than a shutdown. Alibaba's Qwen team appears to be moving its flagship model line toward an API-only distribution model, which means future releases at the top tier may never see public weight drops. The 397B A17B configuration exists today precisely because it was released before that policy change took full effect—but don't count on Qwen4 or beyond following the same pattern.

Why This Matters for Open-Source AI

This isn't just about one model family. When frontier-tier open-source labs start treating their largest, most capable models as API-only products, it fundamentally changes what 'open' means in AI. Smaller researchers can still run 7B and 14B models locally, but the cutting-edge capabilities—the ones that actually advance research—become gated behind pay-per-token access. The community built around open weights assumes these drops will keep coming. That assumption is now being stress-tested.

What You Should Do Right Now

If you've been putting off downloading Qwen3 397B weights, stop waiting. Grab them while they're still available through standard HuggingFace mirrors or official Alibaba repositories. Yes, you'll need significant GPU resources to run it meaningfully—think A100-level hardware for reasonable inference speeds—but the option exists today in a way that may not be true six months from now.

Key Takeaways

  • Qwen3 397B A17B weights ARE publicly available under Apache 2.0—no rumors of shutdown are accurate
  • This appears to be a last-of-its-kind release as flagship models shift toward API-only distribution
  • The move mirrors patterns seen across the industry: keep smaller models open, gate the frontier
  • Local deployment of top-tier weights may become increasingly rare in 2026 and beyond

The Bottom Line

The AI community got lucky with this drop. Don't mistake it for a trend. When you see flagship model families quietly moving their best capabilities behind API walls, that's not an accident—it's a business decision. Download what you can, while you can, because the era of truly open frontier models is ending faster than most people realize.