Weekly AI Rankings — September 06 – September 13, 2026

This week, GigaChat 3.5 Reasoning was released under an MIT license along with artifacts, sparking community discussions. The model is positioned as reasoning-oriented, with quality improvements noted across several benchmarks.

The model features MoE architectural solutions and post-training through online RL with multiple experts, highlighting its reasoning capabilities.

This week, GPT-6 Astra successfully completed the Portal autonomously, drawing attention to its capabilities. Logs indicate a significant number of processed tokens and tool calls.

The agent utilized a cycle of 'screenshot → coordinates and angle → plan → commands', allowing it to complete the task without human intervention.

Discussions this week focused on the practical use of GLM and Qwen, comparing their cost-effectiveness for coding tasks. Additionally, GLM-5.3 was deployed in a Russian cloud.

GLM-5.3 is available via an OpenAI-compatible API, ensuring that user data remains within the country.

GLM-5.3 was deployed in a Russian cloud via an OpenAI-compatible API, generating interest in its availability and pricing policy. Discussions focused on user data retention.

The access price for GLM-5.3 will remain at the level of GLM-5.2, making it accessible to users.

MiniCPM5-2B was released as a compact open-source model for edge and local deployment, drawing attention to its capabilities. Discussions focused on its performance across various benchmarks.

The model features ~2B parameters and includes checkpoints, datasets, and RL stack components for reproducibility.

This week, the hyperframes tool was introduced for creating reproducible generative animation pipelines. It simplifies the integration process into web pipelines.

The tool allows users to describe the steps of motion visual synthesis, simplifying the reproduction of demonstrations.

Discussions this week revolved around GPT-6 Astra's autonomous completion of the Portal, showcasing the agent's capabilities. Logs indicate the efficiency of token processing.

The agent utilized a cycle of actions, allowing it to successfully complete the task without human intervention.

OpenAI released a specialized version of ChatGPT for financial services, generating interest in its functionality. Discussions focused on the choice between various models for specific tasks.

The specialized version includes built-in financial datasets and citation mechanisms, making it suitable for corporate use.

MiniCPM5-2B was released as an open-source model, drawing attention to its capabilities for local deployment. Discussions focused on its performance across benchmarks.

The model includes checkpoints and RL stack components for reproducibility, making it convenient for developers.

Hugging Face introduced ML Intern for the full cycle of ML experiments, generating interest in process automation. Discussions focused on its functionality and applications.

The agent selects models and datasets, estimates computations, and proposes a budget before starting, simplifying ML workflows.