AI News — August 15, 2026: Qwen3.8 27B Runs at 138 tok/s, GLM-5.3 Quietly Stacking CVEs

Qwen released Qwen 3.8 27B, a dense model that runs at approximately 138 tokens/second on an RTX 5090 with the ninfer inference engine, and Z.AI released GLM-5.3, a coding-focused model that achieves Kimi K3-level performance with about a third of the parameters and is being used as an autonomous vulnerability scanner. Apple trained a China-specific AI model with Alibaba, becoming the first US company approved to offer its own model in China, while Google open-sourced HEIR, a compiler for homomorphic encryption, and Meta released Glimmer, an open-weight model, with its more capable Muse Spark model remaining proprietary.

Good morning. It’s a model-release day again — Qwen shipped a 27B that people are running on laptops, Z.AI’s GLM-5.3 is quietly finding CVEs at scale, and the Opus 5 backlash is now loud enough that Anthropic’s phone should be ringing. Meanwhile Google wants you to believe in homomorphic encryption, and Apple did the unthinkable by training a China-specific model with Alibaba. Qwen 3.8 27B lands well. The Qwen team released Qwen 3.8 27B https://huggingface.co/Qwen/Qwen3.8-27B-FP8 , a dense model that HN commenters https://news.ycombinator.com/item?id=49299605 are calling the second local model after Gemma 4 to correctly reason through their private benchmarks. One user reports ~138 tokens/second on an RTX 5090 with the ninfer inference engine, roughly double a naive llama.cpp setup. The main complaints: it overthinks, and its internal monologue has drifted into a strange caveman shorthand “Need be helpful concise”, “Need maybe not overdo” — but the image-to-HTML tests look genuinely close to models several times its size. GLM-5.3 finds CVEs while you sleep. Z.AI released GLM-5.3 https://z.ai/blog/glm-5.3 , a coding-focused model that hits roughly Kimi K3 territory with about a third of the parameters, achieved through post-training scaling alone. The more interesting detail: Z.AI is running it as an autonomous vulnerability scanner across popular OSS at cvd.z.ai https://cvd.z.ai/ , with a stack of critical CVEs currently under embargo. One HN commenter https://news.ycombinator.com/item?id=49294997 upgraded from the $18 to the $80 plan the same day, citing it as the first model that would actually execute a red-team scenario end to end. Opus 5 feels worse, and everyone’s saying so. A widely-shared post https://mun-logadan.github.io/why-does-opus-5-feel-worse/ argues Claude Opus 5 is benchmark-strong but frustrating to work with — it barrels ahead with assumptions instead of asking clarifying questions, and writes in strange orbital prose “sentences that orbit a point, then jump to it like it’s a revealed insight” . The HN thread https://news.ycombinator.com/item?id=49296740 is full of developers reverting to 4.7/4.8 or defecting to OpenAI’s Sol, with one recurring theory: post-training is now optimizing for agent-to-agent handoffs, not humans. As one commenter put it, expect calls from major-corp CEOs before Anthropic responds publicly. Apple ships a China-specific model with Alibaba. Apple has trained a proprietary AI model for the Chinese market in partnership with Alibaba, per The Verge https://www.theverge.com/ai-artificial-intelligence/980160/apple-intelligence-china-custom-ai-model-alibaba , making it the first US company approved to offer its own model in China after clearing the cyberspace regulator. Apple Intelligence is expected to launch in China within months via an iOS update — a notable pivot from its earlier plan to just plug in a domestic third-party model. Google pitches homomorphic encryption for AI. Google open-sourced HEIR https://blog.google/security/how-google-is-making-private-ai-practical-with-homomorphic-encryption/ , a compiler that converts pre-trained models to run on encrypted data, with a demo doing content recommendations without seeing user features. The HN reception https://news.ycombinator.com/item?id=49300314 was rough: the blog post doesn’t mention the ~1000x compute overhead that has kept FHE impractical for years, and one commenter with a master’s thesis in the area said commercial viability is still not close. Another sharp jab: this is the same Google whose password manager still lacks E2EE by default. Meta’s “open” AI, with an asterisk. Meta released Glimmer, an open-weight model, alongside a 6,500-word Zuckerberg letter arguing AI should be “for everyone” — but TechCrunch’s Equity podcast https://techcrunch.com/podcast/metas-open-ai-and-a-250m-deal-gone-very-wrong/ notes the more capable Muse Spark model stays behind proprietary APIs, so “open” is doing some work in that sentence. Elsewhere, French startup Kog claims 30x inference speedups https://techcrunch.com/2026/08/14/kog-is-going-deeper-to-squeeze-more-inference-out-of-gpus/ on existing MI300X and H200 hardware, demoing 3,000 tokens/sec on a 2B model — the open question is whether the technique scales to the larger models enterprises actually deploy. That’s the briefing. If you’re on an RTX 5090, today’s homework is obvious; if you’re on Opus 5, maybe give 4.8 another look.