AI research & open-source LLM model Brief — 2026-09-22

Top Stories

Source: Xiaomi MiMo · September 22, 2026
Summary: Xiaomi officially released and open-sourced the MiMo-V2.6 series, comprising the multimodal MiMo-V2.6-Pro and MiMo-V2.6-Flash. The release emphasizes large-scale reinforcement learning, including six days of live RL training, roughly 750,000 trajectories, and more than 7,000 open RL task environments covering software engineering, vulnerability reproduction, knowledge-intensive work, and web development. Xiaomi also released model weights, technical reports, RL training resources, and an end-to-end training framework.
Why It Matters: The release shifts open-model competition toward RL infrastructure and reproducible agent training, rather than simply increasing parameter counts. Opening the training environments and harnesses could be as strategically important as the model weights for researchers building the next generation of agentic systems.
URL: https://mimo.mi.com/docs/en-US/news/latest/v2-6

Source: China Telecom AI via GlobeNewswire · September 22, 2026
Summary: China Telecom AI released Xing4.0-29B-A4B, a Mixture-of-Experts model with 29 billion total parameters and 4 billion activated parameters. It supports a 256K-token context window, multi-step planning, tool calling, and agentic execution, while the company says low-bit quantization and memory optimization reduce the deployment requirement to about 15GB of GPU memory. The model has been released through GitHub and Hugging Face under an open ecosystem strategy.
Why It Matters: The release illustrates the growing importance of deployment efficiency over raw model scale. A 29B-class agentic model targeting single-GPU operation makes local inference increasingly viable for developers and enterprises that need lower cost, data control, or on-premise deployment.
URL: https://www.streetinsider.com/Globe+Newswire/China+Telecom+AI+Officially+Releases+Xing4.0-29B+Agentic+Large+Model+for+Single-GPU+Deployment/27086594.html

Source: Reuters · September 22, 2026
Summary: Alibaba announced plans for a next-generation AI model with between 5 trillion and 10 trillion parameters, substantially exceeding its current Qwen3.8-Max flagship. At its Apsara conference, the company also unveiled the Zhenwu V900 AI chip, which it says delivers three times the performance of its predecessor and is designed to support large AI clusters. Alibaba is simultaneously targeting more than 20 gigawatts of data-center capacity by 2032.
Why It Matters: The announcement highlights a renewed frontier-scale race in model size, silicon, and infrastructure. For the open-model ecosystem, the development also underscores a widening strategic split between enormous frontier systems and smaller, more efficient models designed for practical local and enterprise deployment.
URL: https://www.reuters.com/business/retail-consumer/alibaba-plans-ai-model-with-5-trillion-10-trillion-parameters-unveils-new-chip-2026-09-22/