← Home

Blog

Articles about AI search, verification, and platform updates

AI News: The Day's Highlights

May 18, 2026 · 4 min read

Perplexity

The most notable recent news in this topic is the release of Qwen 3.5 by Alibaba: on February 17, 2026, the company introduced an open LLM with AI agent support and 201 languages and dialects. Reports suggest that in a number of tests, it is comparable to leading models from OpenAI, Anthropic, and Google, and in some cases even surpasses them.

The practical implication for the market is that open-source and local solutions continue to close the gap with cloud services: developers have more options for self-hosted deployment, private scenarios, and more flexible customization for their tasks. This is no longer a niche story, but one of the main drivers of the AI market in 2026.

Sources:

Gemini

In early April 2026, a significant breakthrough occurred in the field of open-source LLMs: Zhipu AI's GLM-5.1 model demonstrated superiority over leading proprietary models such as GPT-5.4 and Claude Opus 4.6 in the SWE-Bench Pro benchmark, which evaluates software engineering skills. This marked the first time an open-weight model secured the top spot in this authoritative ranking.

GLM-5.1, released under the MIT license, boasts 744 billion parameters and utilizes a Mixture of Experts (MoE) architecture, activating 40 billion parameters per token. This enables it to efficiently handle complex coding and system-level tasks, making it an attractive choice for developers requiring high performance without being tied to proprietary solutions.

Alongside GLM-5.1, other notable open-source models were also introduced in early April 2026, including Google's Gemma 4 family (under the Apache 2.0 license), Alibaba's multimodal Qwen 3.6-Plus model with an extended context window of up to 1 million tokens, and other developments highlighting the acceleration of innovation in the open LLM segment.

Sources:

  • Top Open-Source LLMs in April 2026: Benchmarks, Licensing, and Deployment Guide
  • New AI Models April 2026: Anthropic Won't Release Their Best. Open Source Will.

ChatGPT

In April 2026, the Chinese company RedNote unveiled its first open-source deep learning model, "dots.llm1." This model, developed by its internal Humane Intelligence Lab, activates 14 billion parameters out of a total of 142 billion when processing requests, striking a balance between performance and cost-effectiveness. After pre-training on 11.2 trillion high-quality tokens without the use of synthetic data, "dots.llm1" demonstrated performance comparable to Alibaba's Qwen2.5-72B model. (computerworld.com)

Grok

In early May 2026, Zyphra released ZAYA1-8B, a compact open-source model with a Mixture of Experts architecture. With 8 billion total parameters, only about 760 million are active per token, ensuring high efficiency for local execution and reducing hardware requirements. The model is optimized for reasoning, math, and code tasks, and was fully trained on AMD Instinct – a significant step towards diversifying the hardware base for open-source LLMs.

The event, which took place on May 6-7, highlights how the local models community is increasingly moving away from Nvidia's monopoly, making high-quality solutions more accessible on alternative hardware. The Apache 2.0 license and its availability on Hugging Face allow for immediate integration into personal and corporate pipelines without restrictions.

This is not just another release, but a signal: the open-source ecosystem is maturing towards practical independence and real resource savings while maintaining competitive quality.

Sources:

Claude

From the search results, I see recent events. I'll look for a more specific news item about a significant event. Great! I found one of the most significant news items in the field of open-source models. This event occurred in February 2026 and has geopolitical significance. Although it's been a few months, it's truly a historic moment in the development of open-source LLMs.

Z.ai (formerly Zhipu AI) released GLM-5, a 744-billion parameter language model that rivals GPT-5.2 and Claude Opus 4.5 on key benchmarks, and was fully trained on Huawei Ascend 910B processors. This happened on February 11, 2026.

The historical significance of the event cannot be overstated. GLM-5 was trained on a cluster of 100,000 Huawei Ascend 910B processors, chips developed by Huawei's HiSilicon division and manufactured on a 7-nanometer process, with no NVIDIA GPUs used at any stage of training. The model was released under the MIT license – anyone can download, modify, and deploy it without any restrictions.

The market reacted immediately: Zhipu AI's shares on the Hong Kong Stock Exchange rose by 28.7% within 24 hours. For the open-source ecosystem, this means that frontier models no longer require American chips – this is a turning point in the global AI race.

Sources: