Z.ai confirms Ox Alpha is a new GLM-series model and will release its weights

Commenters spent less time celebrating the GLM lineage confirmation than trying to establish whether Ox Alpha’s claimed capability is real. Benchmark evidence was treated as unstable: LiveBench reportedly places it below a lightweight GPT-5 variant, while an unofficial site briefly showed it beating a stronger rival. Readers flagged the latter as incomplete, sponsor-tainted, and unsuitable for a serious comparison. Speculation that it was distilled from another model also met resistance; several argued that abrupt capability jumps are normal enough that distillation needs hard evidence. Hands-on reports were more encouraging but still uneven. One long-running agentic code-porting task reportedly outclassed several cheaper or faster models, and another reader saw it generate substantial Java bindings in a single huge-context session. Others described a small-looking model that can repair its own mistakes through extra turns and tools, though at very slow inference. Skeptics were unimpressed outside these anecdotes and expect interest to collapse once free access ends. The promised weight release drew approval as a way to strengthen open-weight competition, but readers want to know the model’s size, whether all weights will actually be available, and whether licensing will leave the vendor as the only practical provider. Some also see confusing Chinese-model branding as an adoption handicap, while others say developers already understand the GLM/Z.ai relationship.