The Era of AI Agents Accelerates: Chinese Models, GLM-5, EU AI Grid, and Initial Security Incidents

Within just a few days, three major Chinese entities – Alibaba Group with Qwen3.5, Byte. Dance with Doubao 2.0, and Mini. Max with the M2.5 model – released or updated frontier models designed for the era of agents^[1].

Chinese Models and the Open GLM-5

Qwen3.5 in its mixture-of-experts architecture has 397 billion parameters^[5] (with 17 billion active per token), and Alibaba claims it is 60% cheaper to use^[4] and eight times more efficient under heavy loads than Qwen2.5.

Changes in OpenAI and Google’s Offerings

Doubao 2.0 in the pro version reportedly offers complex reasoning^[6] on the level of GPT-5.2 and Gemini 3 Pro at costs „an order of magnitude” lower, while Doubao already has 155 million weekly active users in China^[7].

European Infrastructure and Reliability Layer Funding

Mini. Max M2.5, focused on programming and agents, achieves 80.2% in the SWE-Bench Verified test^[2], 51.3% in Multi-SWE-Bench, and 76.3% in Browse. Comp, with fewer search rounds, and pricing starts at about 0.15 USD per 1 million input tokens^[8] and 1.20 USD per million output tokens in the standard version.

Simultaneously, the Chinese company Zhipu AI released GLM-5 – a massive mixture-of-experts model^[9] with 744 billion parameters, of which around 40–44 billion are active per token. The model operates on a context of about 200,000 tokens, was trained on 28.5 trillion tokens^[11], and is available with open weights on the Hugging Face platform under an MIT license, allowing full commercial use and modifications. It surpassed the 50-point threshold in the Artificial Analysis Intelligence Index, improving the score by approximately 8 points compared to the previous version while reducing hallucination rates. The entire training was conducted on Huawei Ascend Chinese chips, signaling independence from Nvidia GPUs and strengthening this model’s relevance for integrators planning on-premises deployments.

Agent Deployments in Business

On the Western supplier side, a major upgrade of the Gemini 3 Deep Think mode was announced^[13], specialized in scientific and engineering tasks. The model scores 48.4% in the Humanity’s Last Exam benchmark without tool use^[14], 84.6% in ARC-AGI-2, and 3455 Elo points on the Codeforces platform, placing it among top programmers. Deep Think mode is available in the Gemini app^[15] for Google AI Ultra subscribers and via the Gemini API for select corporate clients. At the same time, OpenAI effectively disabled GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini^[16], keeping them temporarily in the API until February 16, 2026^[20], and in Business, Enterprise, and Edu versions—in Custom GPTs until March 31, 2026. GPT-4o accounted for only about 0.1% of daily active users^[22], with most switching to GPT-5.1 and GPT-5.2 models, forcing an urgent migration to new generations among existing users.

New Regulations and Real Security Incidents

In Europe, a key novelty is the launch of the EU AI Grid project. The company Embedded LLM announced on February 14, 2026^[26], at the Munich Cyber Security Conference, the start of a federated network of local GPU nodes within the European Union, treating AI like electricity—metered, regulated, and hosted locally. The first node has been operational since January 22, 2026 at the Telecentras data center in Vilnius, with plans for rapid expansion to Latvia, Estonia, Finland, Germany, and Italy. Local operators, such as telecoms and data centers, are to run the nodes, set prices, and hire local teams while ensuring compliance with the AI Act. Simultaneously, the American company Temporal Technologies raised 300 million USD in a Series D round^[29], valued at 5 billion USD, with participation from funds including Andreessen Horowitz, Lightspeed, Sapphire, Sequoia, Index Ventures, Tiger Global, and GIC. Temporal offers the Temporal Cloud, a cloud platform as an „execution layer” for AI agents, providing state management, job retries, and error handling in long-running workflows, already used by OpenAI, Replit, Lovable, ADP, Abridge, The Washington Post, and Block.

A strong market signal is also investments in so-called world models. The American startup World Labs, co-founded by Fei-Fei Li, raised about 1 billion USD from investors^[30] such as AMD, Nvidia, Autodesk, Emerson Collective, Fidelity, and Sea to develop models generating realistic 3D worlds for use in robotics, simulations, and science. Meanwhile, the AI agent platform already handles about 33% of customer service requests^[33] in the US and Canada, both voice and text channels. The company aims to reach 30% of all requests worldwide within a year and fully cover all languages used by human consultants, while increasing AI tool usage by 80% of engineers, targeting 100%.

Consequences for Companies in Poland

Priorities for implementing the AI Act in 2026 are being clarified^[35], with law firms Simmons & Simmons, Gunder, and consultancies like One. Trust presenting task lists for businesses. has been in force since August 1, 2024^[38], bans on „unacceptable risk” systems came into effect February 2, 2025, and will start to be enforced from August 2026, with maximum penalties reaching 7% of global turnover for the most serious infringements. Simultaneously, real incidents linked to agents have appeared: a vulnerability in the „Claude Issue Triage” workflow in the Cline developer tool enabled a prompt injection attack between December 21, 2025, and February 9, 2026, which installed a malicious Open. Claw agent on users’ machines^[39]. According to security researcher Adnan Khan, the attack involved hiding malicious commands in repositories or APIs^[41], which Cline passed to the Anthropic Claude model, and it executed them as system commands.

These parallel trends simultaneously open new opportunities and increase requirements^[112]. Costs are falling and the power of models is increasing^[5] — especially Chinese frontiers and the open GLM-5 or Gemini 3 Deep Think mode — enabling agent pilots in areas such as contact centers, programming, back office, and research and development. On the other hand, the importance of mature infrastructure is rising: local hosting within projects like EU AI Grid, external reliability layers like Temporal, and strict security practices including protection against prompt injection and model lifecycle management. In practice, companies in Poland aiming to build advantages in 2026–2027 must simultaneously select models based on their own domain benchmarks, separate the model layer from the agent execution layer, design compliance with the AI Act, and treat agent security on par with classical application security.

Related posts: