This week, Z.AI (Zhipu) confirmed what the community had been speculating about for two weeks: the anonymous model "Ox Alpha" that quietly climbed AI leaderboards to performance rivalling DeepSeek V3.2 is the next generation of their GLM series — and the weights will be open-sourced tonight.
What this is
Over the past two weeks, the AI evaluation community noticed an anonymous model "Ox Alpha" breaking into the global top tier across multiple public leaderboards, on par with DeepSeek. "Stealth models" — anonymous releases used to test real-world capability before formal launch — have become common practice in the frontier AI scene over the past two years: register under a shell name to gauge true performance, then claim credit. Z.AI voluntarily claiming Ox Alpha and releasing weights publicly marks the second Chinese frontier AI firm — after DeepSeek — to explicitly follow the "anonymous leaderboard climb, then open source to capture developers" playbook. "Open source as ecosystem play" is shifting from DeepSeek's signature move to the default operating procedure for Chinese frontier model companies.
Industry view
Optimists see open source as the antidote to overseas restrictions: when GPT and Claude are constrained by export controls, more open weights mean more autonomy and self-control, available for both large domestic players and startups to build on.
But we also hear plenty of measured voices. One pointed question: training frontier models costs run into the tens of billions — how do you monetize after going open source? Z.AI has built up a ToB (enterprise sales) presence over the past few years, but its revenue scale isn't in the same league as DeepSeek's. For Z.AI, open-sourcing may not exactly be "charity" — it's more about telling a compelling story to investors and recruits. Another risk lies in the "rival DeepSeek" narrative: leaderboard scores and real business deployment are often different things. Without differentiated follow-through, this becomes yet another round of leaderboard theatre.
What deserves our attention: after tonight's open-source release, download counts on GitHub and HuggingFace, and the volume of community-driven fine-tunes — those are the hard metrics for judging whether Ox Alpha is a real player or a marketing stunt.
Impact on regular people
Enterprise IT: Another free, commercially usable domestic frontier model option. The cost of self-built AI systems continues to drop, with one more competitor alongside Tongyi and DeepSeek for price comparison.
Individual professionals: The GLM series is already integrated into multiple office collaboration tools. After Ox Alpha goes open source, the ceiling for domestic AI assistants will likely rise another notch — drafting, spreadsheets, and document reading will all feel smoother.
Consumer market: No visible change in the short term — the technology lands in the B-end (enterprise) first. Once the open-source ecosystem grows, the free AI tools available to C-end (consumer) users will multiply and improve.