AI this week: GLM-5.3’s security skills, the push for speed, DeepSeek’s open harness
Z.ai’s GLM-5.3 ties the open leader, OpenAI and Cerebras preview a 750-tokens-per-second tier, and DeepSeek-V4-Pro goes official.
August 28, 2026 · By Bohol AI
Open models got better at security work, speed became a selling point, and DeepSeek released the tool it uses to test its own model.
GLM-5.3
Z.ai’s GLM-5.3 effectively tied Kimi K3, the open-weights leader, on Artificial Analysis’ intelligence index. Z.ai improved it by fine-tuning GLM-5.2 rather than training from scratch, and said the model got good enough at finding and exploiting software vulnerabilities that it ran safety tests before releasing the weights.
AI gets faster
OpenAI and Cerebras previewed Ultrafast, a service tier that runs GPT-5.6 Sol on Cerebras chips at up to 750 output tokens per second. Artificial Analysis measures the same model at about 65 tokens per second on OpenAI’s usual service at maximum reasoning. Google and Nvidia also released fast models this week.
Why it matters: speed decides whether voice assistants and live chat feel natural or awkward.
DeepSeek-V4-Pro goes official
DeepSeek released DeepSeek-V4-Pro-0813, the final version of its larger V4 model, along with the free, open-source harness it used for benchmarking. Its API prices went up.
Research: agents that clean up their own memory
Researchers at Xiaohongshu (RedNote) built a way for AI agents to decide which parts of their working memory to keep, instead of simply dropping the oldest material.
What it means for Bohol
- When choosing an AI tool for customer chat, compare response speed as well as price and quality.
- Free, open tools for testing agents lower the cost of building your own.
Sources
- Z.ai: GLM-5.3
- OpenAI: previewing Ultrafast
- Google: Gemini 3.7 Flash
- DeepSeek: V4-Pro-0813 release notes
- DeepSeek harness
- Hao et al., self-governing context for agents (arXiv)
- The Batch, issue 368 (DeepLearning.AI)