Weekly AI Rankings — August 23 – August 30, 2026

This week, GLM-5.3 garnered attention as an open model for local deployment, particularly in audio content generation. Discussions focused on its open weights and potential use in podcasts and radio shows.

GLM-5.3 features a heavily quantized 2-bit version, allowing it to run on machines with limited resources. The model has also been tested for token efficiency compared to more expensive alternatives.

This week, fal.ai launched a turbo version of Minimax H3 that generates video faster than real-time. Discussions focused on the economics of infinite content generation.

The announcement of MiniMax H3 Max comes with reports of significant speed improvements, reaching approximately 36 times faster generation. The model also features open weights, enhancing its capabilities.

Qwen 3.8-Flash was discussed this week due to its configuration with 125B parameters and long context capabilities. Participants noted its potential in handling large data tasks.

The model activates a limited set of parameters per token, allowing it to efficiently handle long contexts, claimed to be up to 262K tokens.

Discussions around Qwen 3.8 this week focused on recommendations for setting reasoning_effort to optimize performance. Participants shared their experiences using the model.

The Qwen 3.8 model has 27B parameters and requires tuning to avoid excessive reasoning chains.

GigaChat 3.5 Ultra was mentioned this week in connection with its successful completion of a cybersecurity exam, highlighting its capabilities in this area.

The model scored 14% above the passing mark, indicating its high level of preparedness.

This week, Sber announced the GigaEmbeddings lineup, consisting of three models for semantic search and classification tasks. Discussions focused on the open weights and their applications.

The models come in sizes of 480M, 3B, and 10B-A1.8B and are available under the MIT license, allowing for free use.

The GigaEmbeddings lineup from Sber was also discussed in the context of its application in RAG and clustering tasks. Participants noted its openness and accessibility.

The models are designed to convert texts into vectors, making them useful for various AI tasks.

Cursor was at the center of discussions this week due to OpenAI's decision to cut access to its models following the acquisition by SpaceX. This raised concerns about security.

OpenAI explained its decision by citing distrust in compliance with usage terms following the change of control.

MiniMax Design introduced a new agent for video production this week that breaks tasks into steps and utilizes the new H3 model. This sparked interest in automation capabilities.

The application includes tools for scriptwriting, storyboarding, and video generation with voiceovers.

Visual GenAI was discussed as an open course on visual generative AI, offering practical materials and assignments. This highlights the growing interest in education in this field.

The course includes modules on diffusion models and multimodal approaches.