Latest open artifacts (#24): Motif-3, GLM-5.3, Hy4-preview and open model licenses

What it claims

Two things: a licensing trend and a set of model picks. On licensing, Western labs are loosening (Google and Meta moved to Apache 2.0) while frontier Chinese labs are tightening. The new example is Zhipu's GLM-5.3. It moved from MIT, which covered GLM-5.2 and earlier, to a custom licence. The page quotes the operative clause: if the licensee "or any of its affiliates operates a Model as a Service business" and the aggregate revenue of licensee plus affiliates exceeds US$10 billion over any consecutive 12 months, it "must pass Z.AI's security review before using the Software or its derivative works for any commercial purpose." The scope and method of the review are "reasonably determined by Z.AI." The authors state they are not legal experts.

How to use it

Releases covered, with the details the page gives:

Kimi K3's licence requires a commercial agreement for anyone running inference or fine-tuning services. MiniMax M3's licence requires an agreement above a revenue threshold and lists prohibited use cases. The page does not give that threshold.

Limits

• GLM-5.3 "affiliates" is undefined in the English text. The Chinese version uses "关联方", which does have a definition in Chinese law. For a large company, or a subsidiary of one, this ambiguity decides whether the $10B test applies. The English and Chinese texts may not read identically, so check which one governs before deploying GLM-5.3 or its derivatives as a paid service.
• The security review's scope is set unilaterally by Z.AI. No criteria or timeline are given.
• The licence change is version-specific. GLM-5.2 and earlier remain MIT. Pinning to 5.2 avoids the new clause.
• Qwen's licensing is not uniform across the family. The 2.4T flagship uses a custom licence. The page does not state the licence for Flash-Next.
• Hy4-preview overthinks. The authors flag this as a current issue. Their optimism about the final release rests only on how Hy3-preview turned into Hy3.
• Qwen3.8-Flash-Next's GDN / Sparse Attention stack is something the authors expect "the ecosystem to fix integrations" for by the time Qwen4 ships. Expect inference-engine support to lag.
• The dots3 IMO result used an internal harness. It is not a bare-model number.
• The page gives no benchmark numbers for any model. "Impressive scores for its size" (Motif-3) and "competent" (Hy4) are editorial judgments. Per-model details are deferred to the Interconnects Artifacts Hub.
• The GLM-5.3-Flash stealth launch is presented partly as a hype and anti-"benchmaxxing" tactic, not as evidence of quality.