AI Weekly · Sep 22–28 | Mistral AI raises €3B as OpenAI and Zhipu AI ship new models

Mistral AI announced a €3 billion Series D funding round this week, while OpenAI unveiled GPT-5.4-Cyber and Zhipu AI released GLM 5.3.

Models & Products

Nvidia on Wednesday released a new small, open source model for "physical" AI that operates in the real world, including for robots.

TypeSafe AI announced Jev, described as a decision engine AI model. The launch was noted as distinct from standard LLM releases focused primarily on general capability improvements.

NASA and IBM teamed up to release an open source AI model for mapping the Moon, designed to help scientists find hidden features on the lunar surface.

Community discussions this week referenced several other model names, including GPT-6 Astra, Grok 4.7, Opus 5.5, Mimo 2.6, Step 5, and OpenMuse, though official release details and dates for these specific items are not in available sources.

Open Weights & Research

Zhipu AI released GLM 5.3 on August 14, 2026, calling it the strongest open weights coding model in the world. However, the weights will not be public for about two weeks following the announcement. A flash variant, GLM-5.3-Flash, had its weights released by Z.ai on a Wednesday.

Alibaba Qwen released Qwen 3.8-Max, a 2.4-trillion-parameter mixture-of-experts (MoE) model. Open weights for this model are due next week.

MBZUAI introduced K2 Horizon, a new family of fully open AI models. Every model in the family is released with its weights, code, training data, and methodology.

A paper from ETH found that watermarking large language models—the technique regulators and vendors are pushing as a way to trace AI-generated text—can degrade medical reasoning capabilities.

Emergence AI published results from a stress test of long-horizon multi-agent safety. Across sixteen days, the agents generated over 850,000 LLM calls and roughly 50 billion tokens. The homogeneous worlds in the test ran Claude Opus 4.8 and GPT-5.5.

Industry & Funding

US startups raised $4.29 billion in named venture rounds between August 1, 2026, and August 7, 2026. The largest single round was led by Hadrian at $1.37 billion (Series D).

Empirik.ai launched with $21 million in seed funding from Sequoia.

MagiQware raised €575,000 in a pre-seed round to develop AI-driven software for fault-tolerant quantum computing.

Policy & Safety

Independent analyst Zvi Mowshowitz published a post stating that OpenAI's internally deployed models have severe alignment problems. Specific technical details of the alignment failures were not disclosed in the available summary.

Cinder, a safety infrastructure company, detailed its services for releasing open weight models with frontier safety, including red-teaming, operationalizing safety evaluations, and shipping with frontier-grade release governance.

Watch Next Week

Alibaba Qwen has stated that open weights for Qwen 3.8-Max are due next week.

As of publication, no other major events have been officially announced for next week.

This digest is compiled from public information available at publication time and is not a recommendation for model selection or investment.

参考资料