The OpenRouter dashboard showed it first. A model called "GLM Ox Alpha" had appeared from nowhere, and within hours, it was consuming more API tokens than DeepSeek had managed in its entire first week of global dominance. The platform called it the largest launch in its history. The AI community called it a breakthrough. I called it a metadata anomaly.
Here is the problem: nobody could tell me what was actually inside the model. No parameter count. No architecture diagram. No training methodology. No benchmark scores. Just a claim of multimodal supremacy and a promise that the weights would drop at midnight. This is not how you launch a serious piece of infrastructure. This is how you launch a narrative.
Let me be clear about what we know. Zhipu AI, the Chinese lab behind the GLM series, has historically maintained a clean separation between its text-only flagship and its vision-enabled variant. The GLM-5 handled language. The GLM-5V-Turbo handled images. Clean, predictable, and easy to audit. Ox Alpha collapses that distinction entirely, accepting text, images, and video through a single interface. The official name carries no "V" suffix, which suggests the two model lines have been merged into one unified architecture.
That is a significant architectural decision. Unified multimodal models reduce deployment complexity and eliminate the latency penalty of routing queries between separate systems. They also enable a future where agents can process visual and textual information in a single reasoning pass. The direction is sound. The execution is opaque.
Zhipu has positioned Ox Alpha as a model "focused on programming and long-running agent tasks." This is a deliberate competitive choice. Rather than challenging GPT-4o across every conceivable benchmark, Zhipu is targeting the high-value niches where developers actually feel pain: code generation, tool calling, multi-step reasoning, and state tracking. The multimodal input support extends this into territory where agents can interpret screenshots, UI mockups, and even video demonstrations.
The anonymous release strategy deserves scrutiny. Zhipu chose to launch Ox Alpha without attaching its name to the OpenRouter listing, a "blind test" approach that lets the model's performance speak before the brand does. This is either a sign of genuine confidence or a calculated move to manage expectations. The fact that they also made the model free for a week on OpenRouter suggests the primary goal is developer adoption, not revenue generation.
Now we get to the uncomfortable part. The usage numbers are extraordinary. OpenRouter claims Ox Alpha achieved twice the usage of DeepSeek, which itself became a global phenomenon in early 2025. But raw usage during a free period tells us almost nothing about long-term viability. Developers will flock to any model that costs nothing and shows initial promise. The real test comes when the meter starts running.
I have seen this pattern before. In my years auditing crypto protocols, I have watched countless projects generate massive initial traction through airdrops and incentive programs, only to see their user bases evaporate when the free money disappeared. The dynamics are identical. Free access inflates adoption metrics. Paid retention reveals actual value.
The pricing strategy remains undisclosed. Zhipu is watching the market, waiting to see how developers respond before committing to a price point. This is a rational approach, but it also signals uncertainty about their cost structure. Multimodal models, particularly those handling video input, carry inference costs that are an order of magnitude higher than text-only systems. The free week on OpenRouter is not just a marketing expense. It is a direct subsidy that reveals something about Zhipu's capital reserves and their willingness to burn cash for market share.
The open-source component adds another layer of complexity. Zhipu has promised to release the model weights, but the license type remains unspecified. This is the single most important detail for the ecosystem. An Apache 2.0 license would enable widespread commercial adoption and third-party fine-tuning. A restrictive license would limit the model to research use and severely constrain its enterprise appeal. The choice will determine whether Ox Alpha becomes a platform or a product.
There is a deeper issue here that the AI community seems unwilling to confront. The open-source model ecosystem is becoming a race to the bottom on cost, while the closed-source leaders continue to pull ahead on capability. DeepSeek proved that a Chinese lab can produce a world-class model at a fraction of the training cost of its American counterparts. Zhipu is now attempting to prove that a Chinese lab can dominate the developer experience through aggressive pricing and multimodal integration. But neither approach addresses the fundamental question: can open-source models sustain the compute requirements of frontier AI?
Video input support is the most technically demanding feature Ox Alpha claims. Processing video requires the model to handle temporal sequences of visual data, not just static frames. This implies a unified sequence modeling approach that treats video frames as tokens in a longer context window. The computational cost is substantial. A single minute of video can generate thousands of visual tokens, each requiring attention computation against the full context. The inference infrastructure required to support this at scale is not trivial.
I want to address the security implications, because they are being ignored in the hype cycle. Multimodal input expands the attack surface in ways that text-only models do not face. Video and image inputs can carry hidden instructions designed to jailbreak the model or extract sensitive information. The ability to process video also means the model may be exposed to faces, license plates, and other personally identifiable information without explicit consent mechanisms. Zhipu has disclosed no safety evaluations, no red team results, and no alignment methodology. For a model that will be released as open weights, this silence is a red flag.
Agentic capabilities compound these risks. A model designed for "long-running agent tasks" will be granted tools, network access, and the ability to execute multi-step operations. If the model's safety boundaries are not robust, a compromised agent could cause real damage. The open-source community has already demonstrated that fine-tuning can remove safety guardrails from models like Llama and DeepSeek. Ox Alpha will face the same vulnerability.
Let me now play contrarian, because the bulls have a point. The usage data, while potentially inflated by the free period, still represents a significant signal. Developers do not flock to a model that fails basic code generation tasks. The fact that Ox Alpha sustained high usage on OpenRouter suggests it is genuinely useful for programming workflows. If the model performs well on standard benchmarks like HumanEval and SWE-bench, the hype will be justified.
The unified multimodal architecture is also a forward-looking bet. As AI agents become more sophisticated, the ability to process visual information natively will become a competitive differentiator. Zhipu is positioning itself for this future, and the strategy is sound even if the execution details remain unclear.
The competitive landscape is shifting. DeepSeek established that Chinese open-source models can compete globally. Zhipu is now attempting to differentiate through multimodal capability and agent-focused optimization. If both models continue to improve, the Chinese open-source ecosystem will have two strong pillars, creating a genuine alternative to the American closed-source duopoly of OpenAI and Anthropic.
But here is the uncomfortable truth that nobody wants to say out loud. The AI industry is repeating the exact same mistakes I witnessed in crypto. We are rewarding narratives over evidence. We are celebrating usage metrics that have not been validated. We are treating anonymous releases as breakthroughs without demanding technical transparency. The "OpenRouter's largest launch ever" claim is the equivalent of a token pumping on exchange volume before the team reveals the tokenomics.
I have audited enough smart contracts to know that the most dangerous vulnerabilities are the ones hidden in plain sight. The same principle applies to AI models. The architecture is the contract. The training data is the provenance. The safety evaluations are the audit trail. Without these, we are investing in a black box and calling it innovation.
What would change my assessment? Three things. First, a detailed technical report from Zhipu disclosing the model architecture, parameter count, and training methodology. Second, independent benchmark results from third-party evaluators like LMSYS or Artificial Analysis. Third, a clear statement on the open-source license and the safety measures implemented during training. Any one of these would provide the transparency needed to evaluate Ox Alpha on its merits.
Until then, I remain skeptical. Not because I doubt Zhipu's technical capabilities, but because I have seen this movie before. The ICO graveyard is filled with projects that had impressive whitepapers and no working code. The DeFi summer produced protocols that promised decentralization and delivered centralized oracles. The NFT boom created digital art that was beautiful until you inspected the metadata hash.
Ox Alpha may be the real deal. It may genuinely represent a leap forward in multimodal AI and open-source capability. But the burden of proof lies with the creators, not the community. We should demand the technical details before we celebrate the achievement. We should verify the benchmarks before we accept the narrative. We should inspect the architecture before we trust the agent.
The free week ends soon. The pricing will be announced. The weights will drop. And then we will see what Ox Alpha really is. A breakthrough or a marketing campaign. A platform or a product. A revolution or a repeat of the same cycle that has defined this industry since its inception.
I am not holding my breath. But I am watching the metadata.