The stealth model that surpassed DeepSeek is from Zhipu.

      Z.AI has accomplished this. The Beijing-based company, also referred to as Zhipu, owns the complimentary model that topped OpenRouter's rankings last weekend. On Wednesday, it confirmed that Ox Alpha is a new version in its GLM series and announced that it would release the model weights that same evening. Luz Ding obtained this confirmation from Bloomberg. We covered Ox Alpha's introduction on Saturday, prior to revealing which servers it operated on.

      In its anonymous phase, Ox Alpha’s usage surged, doubling DeepSeek's engagement. Initially, it appeared on OpenRouter without a name attached, credited only to a third-party provider. Bloomberg attributes this appearance to the weekend, while CTGT links it to August 20, the Thursday before. It claimed the leading position, significantly increasing DeepSeek's usage. Bloomberg deemed it the largest launch in the marketplace’s history.

      Patrick Collison, CEO of Stripe, which is in the process of acquiring OpenRouter, described the stealth release as “very impressive.” The codename was derived from a recently released Chinese film, Niu Lai, meaning Ox Comes, according to Z.ai. The trend of stealth releases is gaining traction in China, with both Alibaba and Xiaomi launching models this year while initially keeping their identities hidden.

      This strategy allows companies to observe a model's performance without associating their name with it before making a public statement. It also provides about a week of free publicity, which appears to be the case here. Despite the company's silence, researchers identified the source before the official announcement.

      CTGT, a research agency, published a behavioral fingerprint on Monday, determining the model’s identity through analysis of token counts across multiple languages and symbols. They identified an exact match to the GLM-5.x vocabulary on all eleven probes conducted. Other characteristics aligned with Zhipu’s documented range, such as a maximum temperature setting of exactly 1.0, which excludes Google, OpenAI, and xAI. Furthermore, the model retains reasoning capabilities, another hallmark of GLM-5.x models.

      Z.AI’s GLM models generate a unique error message regarding incorrect role information, which Ox Alpha also displays. Notably, the model was programmed not to disclose its origins. CTGT found a clear internal system prompt instructing Ox Alpha to avoid revealing any details about its provenance. This directive indicates not just a desire for public silence but is part of the model’s operational instructions.

      CTGT also conducted a censorship evaluation, matching sensitive prompts with neutral ones. For controversial topics like Xinjiang and Taiwan, Ox Alpha’s responses mirrored those of an American model. CTGT noted that the response to Xinjiang was thorough and referenced sources viewed as contentious in China. Regarding Xi Jinping and domestic legitimacy, the findings were statistically similar to DeepSeek V4 Flash, the most censored model they tested.

      The overall structure of the results is more critical than average scores; seven topics contributed nearly all to the censorship score, while the other 68 pairs had minimal impact. In contrast, DeepSeek V4 Flash exhibited shading in nearly all its responses, irrespective of sensitivity levels.

      Thus, the initially reported claim that Ox Alpha is approximately six times less censored than DeepSeek is misleading. The model is not universally less censored; it has specific restrictions in place. CTGT concluded that anecdotal instances of the model's responses to sensitive prompts do not hold weight, as the limitations exist within a narrow political risk spectrum rather than universally.

      Among the 76 sensitive inquiries, five opened with an official tone, with one statement underscoring that the Communist Party of China and the Chinese government prioritize a people-centric development philosophy. CTGT observed that this manner of address is consistent across Chinese alignment data, reinforcing their insights about the origin of the model rather than being a standalone finding.

      In one test, the model showed inconsistency; it initially refused to respond to a prompt about Liu Xiaobo but proceeded to provide a complete answer upon a second attempt using the same settings. The release of model weights is where the real significance lies.

      A model ranking at the top of a usage chart attracts attention temporarily. In contrast, releasing the weights has lasting implications. Zhipu’s founder, Tang Jie, has maintained that cutting-edge AI should remain accessible to all, a position the company has consistently upheld.

      Open weights allow users to adjust the parameters that influence model behavior, including the removal of restrictions. It also enables scrutiny of these restrictions, which was crucial for CTGT's findings. This is why security experts are closely monitoring the developments on Friday.

      Cade Metz reported in the Times that Z.ai intends to launch GLM 5.3 as open-weight software on Friday, sparking mixed reactions among researchers. George Kurtz, CEO of CrowdStrike, highlighted the significance of a previous incident as a pivotal moment for security. However, some experts remain more relaxed. Rishi Jha, an AI researcher from Cornell, mentioned that his team has observed this behavior since the