Two things happened this week that, together, change how you should think about open-weight models. First, Alibaba dropped the biggest open-weight model by parameter count ever released — Qwen3.8-2.4T-A95B, a 2.4T-parameter MoE with 95B active — but under a custom license that is emphatically not Apache 2.0. Second, the 27B dense variant that most developers will actually run is expected tomorrow, Aug 15. Here's what's real, what's marketing, and what the license actually changes.
1. The release: biggest open weights ever, with a moat attached
Qwen3.8-2.4T-A95B went up on Hugging Face and ModelScope on Aug 12 (HN discussion hit ~700 points within a day). Some quick facts:
- ~4.9TB BF16, ~5TB total — by parameter count, the largest open-weight release to date. The 1-bit quant comes in at a stunning ~397GB with 95B active, which per early testing puts Opus 4.5-class performance inside a machine a normal person could buy.
- It's a Kimi K3 rival in intent — same playbook Moonshot ran in July: open the weights, drive ecosystem adoption, keep the enterprise tier under control.
- Community caveats immediately surfaced: no vision input and no 1M context on the open-weight version (those are Qwen3.8-Max API features), and only BF16/FP8 at launch — no QAT'd q4, so serving it cheaply needs a deep-pocketed quantizer.
2. The license: Apache 2.0 is over for frontier Qwen
This is the part that matters more than any benchmark. The Qwen3.8-Max License replaces the permissive Apache 2.0 used for previous Qwen generations with a custom commercial gate:
- Any MaaS or AI work-assistant business with aggregate revenue above $50M in any rolling 12-month period must negotiate a separate commercial license.
- "MaaS" is defined broadly — any third-party access to inference or fine-tuning where the provider controls inputs or parameters.
- Under $50M, it's free. Over it, you're at the negotiating table with Alibaba.
Read the strategy plainly: platform capture, not charity. The safe harbor keeps startups hooked; the gate forces anyone who scales into a competitor to Alibaba's own API business to pay or fork. Forkast's framing is correct — this is Alibaba drawing a hard line around the commercial value of frontier-class weights, and the market noticed: Alibaba shares jumped ~7% in HK and ~4.5% on the NYSE after the Aug 3 Max announcement.
3. The benchmarks: treat them as marketing
Alibaba's official blog positions Qwen3.8-Max as top-tier — Terminal Bench 2.1 86.6 vs Opus 4.8's 84.6 and GPT-5.6 Sol's 88.8; PaperBench 93.0 vs 90.5. But the comparison is cross-harness and fundamentally apples-to-oranges:
- Qwen was evaluated on OpenCode; Claude on Claude Code (avg@10, 5-hour timeout); GPT-5.6 on Codex.
- Different timeouts, token limits, and harness tooling make those numbers non-comparable. Treat vendor-reported leaderboard deltas of 2-3 points as noise until independent harnesses replicate them.
4. What to actually watch tomorrow
The 27B dense variant (expected Aug 15) is the community-adoption play — the model most people can run locally, in the lineage of Qwen3.6-27B which HN crowned "the sweet spot for local development" (1,192 pts in June). Two questions to answer once it lands:
- Does the 27B keep a permissive license, or does the $50M gate extend down-market? The Max license gates MaaS; if the 27B inherits the same terms, the "free local model" era has a ceiling everywhere, not just on the flagship.
- How does it actually do on SWE-bench Pro with a single harness? Qwen3.8-Max posts 67.7 vs Claude's 69.2 — a real gap that matters for agentic coding, and the 27B will likely land below that.
The bottom line
The open-weights era isn't ending — it's getting a business model. Alibaba is executing the Kimi K3 playbook at $300B+ scale: permissive enough to capture developers, gated enough to tax anyone who builds a competing infrastructure layer. For builders, the new due-diligence question isn't "is the model good" — it's "what happens to my business at $50M revenue." Tomorrow's 27B release will show whether that gate extends to the models most of us actually run.
Top comments (0)