# AI Week: qwen's open weights landed, and the licence is the part that matters > Issue 5. Qwen published two open-weight checkpoints (27B under Apache 2.0, 2.4T under a custom licence), OpenAI wrote up the controls behind its Astra pause, and GLM-5.3 got a primary source but its weights are still out. The hard number: OpenAI puts monitoring overhead at roughly 20 percent of the inference compute it watches. - Author: Sailesh Dahal (https://saileshdahal.com.np) - Published: 2026-08-23 - Reading time: 10 min - Canonical: https://saileshdahal.com.np/ai-week-2026-08-23 - Tags: AI, llm, Security - Series: AI Week (part 5 of 5) at https://saileshdahal.com.np/series/ai-week - Previous: https://saileshdahal.com.np/ai-week-2026-08-17.md --- Two weeks ago I wrote that Qwen had promised to open-weigh a Max-class model, its first, and that as of 9 August no repository existed on Hugging Face. That changed this week. Both checkpoints are up now, and the part worth your attention is not the parameter count. It is that "open weights" means two different licences and two different feature sets in this one release. ## 1. Qwen's open weights landed, and the licence split is the real story In Issue 3 I noted Qwen had promised to open-weigh a Max-class model and that, as of 9 August, no repository existed on Hugging Face to check. Both checkpoints are published now: Qwen3.8-27B and Qwen3.8-2.4T-A95B. The finding is the split between them, not the size of either. Qwen3.8-27B is the straightforward one. It is a 27-billion-parameter dense model under Apache 2.0, with native image and video understanding, 262,144 tokens of native context extendable to 1,000,000, and a quantized build the model card and Unsloth put at about 17 GB of RAM at Q4_K_M. You can drop it into vLLM, SGLang or llama.cpp and treat it like any other permissively licensed open model. Qwen3.8-2.4T-A95B is the one the announcement led with: 2.4 trillion total parameters with 95 billion active, the first Max-class Qwen with published weights. Read the licence before you plan around it, because the open checkpoint is not the hosted API. It ships under a custom Qwen3.8-Max Licence, not Apache 2.0. It is text only, with no vision and without the 1M context window the API advertises, and the repository discussion notes it is roughly 2.5 TB at FP8. So "open at Max scale" overstates what you actually receive. You can study and self-host it, but the terms are not the permissive Apache grant the 27B carries, and serving 2.4T is a datacentre problem most teams will not take on. The model card and the blog are Alibaba's own numbers, and the capability claims (Terminal Bench, SWE-bench Pro, and the rest) are vendor-reported. The structural facts, licence and context window, are checkable on the repository itself, and those are the ones that should drive your decision. ![Qwen 3.8 shipped two open-weight checkpoints in August 2026. Qwen3.8-27B is a 27-billion-parameter dense model under Apache 2.0 with native image and video understanding, 262,144-token native context extendable to 1,000,000, and a roughly 17 GB Q4_K_M quant. Qwen3.8-2.4T-A95B is 2.4 trillion total parameters with 95 billion active under a custom Qwen3.8-Max Licence that is not Apache, is text-only with no 1M context, and needs about 2.5 TB at FP8. The two releases are not interchangeable.](https://saileshdahal.com.np/images/posts/ai-week-2026-08-23/qwen3.8-weights.svg) ## 2. OpenAI wrote up the controls behind the Astra pause In Issue 3 I covered OpenAI's decision to slow Astra, an unreleased model it could not rule out as Critical for cyber capability. This week OpenAI published the operational follow-through, a post dated 18 August titled "Pacing model development in an era of cyber-critical capabilities". It is the most concrete account yet of what the safeguards actually cost. After the OpenAI-Hugging Face incident, the company paused frontier model inference in research clusters for any run that could execute code or reach the internet, then restored a narrower secure path and worked through workloads one by one. It also put a two-week pause on reinforcement-learning training for its latest deployment-intended models, and its largest planned frontier RL run remains on hold while it runs smaller training and evaluation first. On monitoring, OpenAI says overhead is roughly 20 percent of the inference compute being watched, with activation classifiers at every sampled token escalating to automated investigators, and a rule that if a critical-boundary alert cannot be cleared within 30 minutes the activity is paused. Those are OpenAI's own figures, and it says a technical report on the Hugging Face incident is coming but not yet published. The reason I put this at the top of the security thread rather than another breach is that it is the first time a lab has put a number on the tax of its own safeguards. If you build agents, a 20 percent monitoring overhead and a 30-minute containment clock are the shape of the problem you inherit from the people training the models you call. ![OpenAI's 18 August 2026 pacing post describes the controls added after the Hugging Face incident and the Astra pause: a two-week pause in reinforcement-learning training on its latest deployment-intended models, its largest planned frontier RL run left on hold, a pause on frontier inference that could run code or reach the internet, monitoring estimated at roughly 20 percent of the inference compute it watches, and a rule that activity is paused if a critical-boundary alert cannot be cleared within 30 minutes.](https://saileshdahal.com.np/images/posts/ai-week-2026-08-23/openai-pacing.svg) ## 3. GLM-5.3 now has a primary source, but its weights are still out In Issue 4 I left GLM-5.3 out because there was no primary release to inspect, only secondary coverage. There is one now: Z.ai published a blog post on 14 August. So I owe the correction, and the post changes the picture in one way and not another. What is now checkable: GLM-5.3 uses the same base model as GLM-5.2, and Z.ai says every gain comes from post-training, not a new architecture. The 743-billion-parameter count for that base is from secondary coverage, not the Z.ai post I read. The coding numbers are large. Terminal Bench 3.0 goes from 4.6 to 28.3, DeepSWE from 46.2 to 66.9, and Z.ai claims it is the most capable open-weights coding model. It also developed what Z.ai calls emergent cyber capability, and the company has opened a public disclosure ledger, cvd.z.ai, reporting 2,436 vulnerabilities found across 269 projects, 1,097 of them critical or high severity, some flaws decades old. What has not changed: the weights are still not published. The blog says they will land about two weeks after the 14 August launch, after safety evaluation, which puts them around late August. As of today the model is API-only, so none of those benchmark numbers can be independently reproduced yet, and they remain Z.ai's own figures from its own harnesses. Treat the 50 percent coding-improvement claim as a vendor claim, not a measured one, until the weights are out and someone outside Z.ai runs the suite. The disclosure ledger is the part I would actually open. A public, updateable record of real vulnerabilities with severity and CVE where available is more useful to a builder than a benchmark bar, because it is the kind of artifact you can verify row by row. ## 4. OpenAI updated the Model Spec, and o3 leaves ChatGPT this week A quieter change with a longer tail: OpenAI updated its Model Spec on 18 August. The spec is the document that describes intended model behavior, and if you build behavior that depends on it, a spec change can shift outputs you were relying on. This revision adds clarity on appropriate relational interactions for teens, clarifies how assistants should handle false or unsupported premises, removes outdated guidance written for pre-reasoning models, and adds a new section, "Be clear about capabilities and limits". The last of those is the one to note if you trust a model to report what it can and cannot do. The full spec lives at model-spec.openai.com. Separately, OpenAI's release notes confirm o3 retires from ChatGPT on 26 August, the end of a 90-day sunset, with no API changes. If you have a product that sends users into ChatGPT on o3, or you have tuned prompts against o3's behavior there, this is the week to move them. The API path is unaffected, so backend integrations stay put. ## 5. The story I am not carrying forward The loudest AI coverage this week was financial. Bloomberg led 17 August with Anthropic's revenue boom lifting AI stocks, and the funding rounds kept coming, another nine-figure raise most days. None of it changes what you build on Monday. A share price move tells you about investor sentiment, not about whether your retrieval pipeline should switch models, and a funding announcement tells you about a balance sheet, not about an API contract. I would rather keep this issue on the things that alter a deployment: what you can self-host, what the safeguards cost, and what the models can and cannot yet prove. One thread I am still holding, not dropping: Meta's promised retrospective on the Muse Spark breach has not appeared as of 23 August. Issues 3 and 4 both noted it was promised "once we have all the facts". That gap stands, and it is the one I would most like to close next week, because three of the four breach disclosures this month traced to the same evaluation vendor and Meta is the only lab with no document of its own. ## 6. What I would do on Monday Read the Qwen3.8-2.4T-A95B licence before you plan around it. The 27B is your default open option; the 2.4T is a different legal and operational object, text-only and heavy. If you only need a drop-in open model this quarter, the 27B is the one to benchmark, not the headline number. Stand up a local or self-hosted lane with the 27B on the short list. Native vision, 262K context, and an Apache grant make it a credible default for code and document work that should stay inside your boundary. Measure the same ten tasks you would send to a hosted model, and record retries and review time, not just token cost. Treat model-behavior specs as versioned dependencies. OpenAI's Model Spec changed this week; if your prompts or evals assume a specific behavior around capabilities and limits, pin the spec version you tested against and re-run your eval when it moves. Move any ChatGPT-o3 users before 26 August. The API is fine, but the ChatGPT surface is not, and a user who lands on a retired model mid-task is a support ticket you can prevent today. Assume 20 percent is the floor for oversight cost on capable agents. OpenAI's own number is for its own frontier training, but the shape holds: monitoring that can interrupt, not a log you read later, is what the rest of the industry is converging on. Budget for it before you ship an agent that can act. ## Sources Primary: - [Qwen3.8-27B model card, Hugging Face](https://huggingface.co/Qwen/Qwen3.8-27B) - [Qwen3.8-2.4T-A95B model card, Hugging Face](https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B) - [Qwen3.8-Max: A New Bar for Coding and Cowork, Qwen Team](https://qwen.ai/blog?id=qwen3.8) - [Pacing model development in an era of cyber-critical capabilities, OpenAI](https://openai.com/index/pacing-model-development-cyber-capabilities/) - [GLM-5.3: Frontier Coding with Emergent Cyber Capabilities, Z.ai](https://z.ai/blog/glm-5.3) - [Z.ai Security Disclosure Ledger, cvd.z.ai](https://cvd.z.ai/) - [Updates to the OpenAI Model Spec (August 18, 2026), OpenAI Help Center](https://help.openai.com/en/articles/9624314-model-release-notes) - [Model Spec, OpenAI](http://model-spec.openai.com/) Reporting and analysis: - [Anthropic revenue boom boosts AI stocks, Bloomberg, 17 August 2026](https://www.bloomberg.com/news/videos/2026-08-17/anthropic-revenue-boom-boosts-ai-stocks) - [Qwen 3.8 open-weights guide, Codersera](https://codersera.com/blog/qwen-3-5-complete-guide-2026/) - [GLM-5.3 coverage, Eigent AI](https://www.eigent.ai/blog/glm-5-3-coding-cyber-model) - [GLM-5.3 benchmarks, Emergent](https://emergent.sh/learn/glm-5-3-benchmarks) The cover is an original abstract image generated procedurally in greyscale with no text, made for this issue. The two diagrams are original hand-drawn SVG redrawings of the linked primary sources: the Qwen weights diagram is built from the two Hugging Face model cards and the Qwen blog, and the OpenAI pacing diagram is built from the 18 August OpenAI post. Parameter counts, licence terms, context windows and the roughly 20 percent monitoring figure are taken from those sources and remain vendor-reported where noted. If I got something wrong, tell me and I will correct it in the next issue.