--- title: "Qwen, DeepSeek, GLM-5.3, Ox Alpha: the week "open" stopped meaning just one thing" description: "Four releases in the same week of August 2026 — Qwen3.8-Max, DeepSeek V4 Pro, GLM-5.3 and the mysterious Ox Alpha — show four very different degrees of openness. What to check before choosing one for your company." date: 2026-08-24 tags: [Open models, News, Licenses] url: "https://ai.malagoli.me/en/blog/qwen-deepseek-glm-ox-alpha-agosto-2026" locale: en_US image: "/blog/qwen-deepseek-glm-ox-alpha-agosto-2026.svg" --- # Qwen, DeepSeek, GLM-5.3, Ox Alpha: the week "open" stopped meaning just one thing In a single week: MIT weights already downloadable, a freshly published revenue-share license, a model that finds critical vulnerabilities and keeps its weights locked, and a fourth model that doesn't even have a name. Here's what changes for anyone deciding today. *** Four releases within the same handful of days in August 2026 tell the story better than any leaderboard: the word "open" no longer describes a single condition. On August 12, DeepSeek published the full weights of V4-Pro-0813 under the MIT license. The same day, Alibaba released the weights of Qwen3.8-Max — but under a purpose-written license with revenue thresholds that neither Apache nor MIT ever had. On August 14, Zhipu/Z.ai announced GLM-5.3, a model capable of finding more than a thousand critical vulnerabilities in open source software, while withholding its weights for a safety review. And on August 20, a fourth model, Ox Alpha, appeared on OpenRouter with no name, no public license, and data-retention policies that contradict each other depending on which document you read. Lined up together, these four cases form a near-perfect spectrum — and they're a more useful practical test than any benchmark for understanding what to check before choosing a model. ## DeepSeek V4-Pro-0813: the fully open end of the spectrum On August 13, DeepSeek moved V4 Pro out of preview after roughly four months, publishing the full weights of V4-Pro-0813 on Hugging Face: a Mixture-of-Experts model with about 1.6 trillion total parameters and 49 billion active per token, roughly 893 GB across 67 safetensors files, with context up to 1 million tokens and DSpark speculative decoding. The license is MIT, the same already seen on V4-Flash: no revenue thresholds, no geographic restrictions, no conditions beyond the minimal attribution the license itself requires. It's the easiest of the four releases to evaluate: if you need the weights in production, you can download them today and know exactly what you're allowed to do with them. There is one detail worth noting for anyone considering the API instead of self-hosting: from August 16, V4-Pro's API pricing rose to $1.32 per million input tokens and $3.96 per million output tokens, roughly fourteen times the family's cheapest tier (V4-Flash). That's exactly the kind of cost jump that makes self-hosting more attractive as request volume grows — the MIT license makes it a concrete option, not just a theoretical one. ## Qwen3.8-Max: the weights arrived, but read them carefully Two weeks ago we covered how Alibaba had announced Qwen3.8-Max with no weights or license published on launch day. On August 12 that promise was only partly kept: the open checkpoint's weights, published as Qwen3.8-2.4T-A95B (2.4 trillion total parameters, 95 billion active, native context of about 262,000 tokens extendable to roughly 1 million), are now on Hugging Face together with an FP8 variant — but it's a stripped-down version compared with the full Qwen3.8-Max: text only, no vision input, no 1-million-token context active by default. The most notable surprise is in the license: it's not Apache 2.0 like most of the Qwen family, but a purpose-written text, the "Qwen3.8-Max License". It broadly grants use, copying, modification, distribution, sale, hosting and fine-tuning — but adds two conditions Apache 2.0 has never had: anyone exceeding 100 million monthly active users or $20 million in monthly revenue must prominently display the model's name in the product; and anyone operating a Model-as-a-Service business or an "AI Work Assistant" with aggregate revenue above $50 million in any trailing 12-month period must obtain a separate license from Qwen before using the software. For most companies evaluating an internal use case, these thresholds stay theoretical — but it's exactly the kind of clause to read in full before building a commercial product on top of it, not after crossing them. ## GLM-5.3: the model that finds vulnerabilities, with weights still under lock On August 14, Zhipu (Z.ai) introduced GLM-5.3 as "built to code, ready for cyber defense", reusing GLM-5.2's roughly 744-billion-parameter base and achieving its gains through more extensive post-training rather than a new pretraining run. On the in-house Terminal-Bench 3.0 benchmark, the score jumped from 4.6 to 28.3 — a 6.2x improvement, the highest among open models. API access went live on August 18 through the GLM Coding Plan and ZCode. The most unusual detail, though, is what the model did during Z.ai's internal testing: post-training produced, according to the company without this being a stated goal, an exploit-chain reasoning capability that identified 1,097 critical vulnerabilities in Linux, WebKit and FreeBSD, and helped uncover 2,436 vulnerabilities across 269 open source projects — some dating back as far as 45 years. That's part of why Z.ai decided to withhold the open weights, promised in stages after what it calls the most extensive safety review it has ever run on its models, with an indicative target around August 28. As of this writing, GLM-5.3's weights are not yet downloadable: it's accessible only via API. It's a case worth watching more for the precedent it sets than for the model itself: a lab voluntarily delaying an open release because the model turned out more capable than expected in a dual-use domain is a signal that technical capability alone no longer decides when a release is ready. ## Ox Alpha: the fourth model doesn't even have a name On August 20, a model labeled only "Stealth" appeared on OpenRouter and OpenCode, with no official name, press release or declared company — the community nicknamed it Ox Alpha. It offers a 1,048,576-token context window, accepts text, images and video, and is free for roughly a week during the preview. An informal test on ten software-engineering tasks (DeepSWE) credited it with over 80% success, against roughly 65% and 52% for two established competing models — a viral figure but not verified by an independent leaderboard, to be treated with caution. Community fingerprinting points to a possible relationship with Z.ai's GLM family, but no official source has confirmed it. What actually matters for a company, though, isn't who is behind it: it's what happens to the data you send it. Ox Alpha's OpenRouter listing states prompts and completions are retained but not used for training; the End User License Agreement governing all "Stealth" models on the platform instead grants OpenRouter and the anonymous provider the right to use content for training, evaluation and improvement; OpenCode's documentation, a third source, claims a zero-retention policy with no training use at all. Three documents, three different versions — and none of them lets you know under which jurisdiction, or by which company, your data is actually being processed. There's no published license, no downloadable weights: it's the exact opposite of DeepSeek's MIT. ## What to check, in practice, before choosing any of these four models - Are the weights already downloadable today, or only promised? DeepSeek and Qwen are downloadable now; GLM-5.3 is not yet; Ox Alpha, probably never in this form. - Read the license published with this specific release, not the previous one: the Qwen3.8-Max License introduces revenue thresholds Apache 2.0 never had. - A model capable of finding hundreds of critical vulnerabilities deserves more scrutiny before you authorize it on proprietary code, not less — however good it scores on benchmarks. - If you can't name the company operating the model, don't send it trade secrets, credentials, customer repositories, or personal, health or financial data: without a name there's no contract, and without a contract there's no accountability to enforce. This is the check we run every week for our clients before authorizing a model in production: not stopping at the lab's name or its leaderboard score, but reading the license published that specific day, verifying whether the weights are actually downloadable, and understanding who processes your data when you don't keep it on your own infrastructure — because the difference between an open release and an anonymous endpoint, this week more than ever, doesn't show up on the benchmark podium. ### Sources - [Hugging Face — Qwen3.8-2.4T-A95B model card](https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B) - [Hugging Face — DeepSeek-V4-Pro-0813 repository and license file](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813/blob/main/LICENSE) - [Z.ai — official GLM-5.3 announcement](https://z.ai/blog/glm-5.3) - [MarkTechPost — Z.ai ships GLM-5.3 without retraining the base model](https://www.marktechpost.com/2026/08/14/z-ai-ships-glm-5-3-without-retraining-the-base-model-better-at-complex-coding-and-long-horizon-tasks/) - [Tech Times — GLM-5.3's post-training produced unplanned exploit chains, finding 1,097 critical bugs](https://www.techtimes.com/articles/324426/20260814/glm-53-post-training-produced-exploit-chains-zai-never-planned-finds-1097-critical-bugs.htm) - [SiliconANGLE — nobody knows who built Ox Alpha or where the code it's fed goes](https://siliconangle.com/2026/08/23/nobody-knows-who-built-ai-coding-model-ox-alpha-or-where-the-code-goes/) - [Tech Times — Ox Alpha retains every prompt: you cannot name the company holding them](https://www.techtimes.com/articles/325244/20260823/coding-model-ox-alpha-retains-every-prompt-you-cannot-name-company-holding-them.htm) *** - [all articles](/en/markdown.md) - [formatted version](/en/blog/qwen-deepseek-glm-ox-alpha-agosto-2026) © 2026 ai.malagoli.me · data hosted in Switzerland · CH ✓ [Privacy Policy](/en/privacy) · [Cookie Policy](/en/cookie-policy)