Weekly AI Rankings — August 02 – August 09, 2026

This week, Qwen 3.8-Max was discussed as Alibaba's new model aimed at autonomous development, featuring 2.4 trillion parameters. The model showcased its capabilities in creating interactive HTML scenes with physics, surpassing competitor Fable 5.

Qwen 3.8-Max can plan tasks, run tools, and correct errors, while also natively handling images. It was tested for scene stability and launch cost efficiency.

Qwen 3.8 was also discussed this week in relation to its integration with Qwen 3.8-Max and its capabilities in voice interaction. The model demonstrates a bidirectional voice stream, although some configurations experience latency issues.

The model includes ASR and TTS components, allowing it to perform background tasks while engaging in interactive communication. An API for Qwen 3.8-Max has also become available.

This week, the Kandinsky WM 1.0 model was introduced by Russian developers for synthesizing physically plausible scenes. It aims to train robots and drones in challenging scenarios.

Kandinsky WM 1.0 is a free neural network that can be used to create world models. It is designed for use in rare or hazardous conditions.

This week, MiniMax released the weights of its video model H3, providing the community access to the model's artifacts. This event followed an announcement and sparked interest among developers.

The H3 model is omni-modal and supports open weights, allowing developers to use it in various applications. Part of the production pipeline remains closed.

DeepSeek V4 Flash 0731 was discussed this week in the context of a practical benchmark showing a performance of around 58 tokens per second. This value has become a benchmark for agentic coding.

The model operates in FP8 mode and was tested on two Spark accelerators. The testing results are used to evaluate throughput.

ChatTJB became a viral case this week as its mechanism, based on human manual responses, drew attention as an example of fake AI services. Registration was suspended due to high demand.

The project features a bot interface, but the actual inference is performed by a real person. This sparked discussions about transparency in AI services.

ChatGPT updated this week, providing free-tier users with unlimited text chats and a new 'Think' button for complex queries. This change sparked discussions about accessibility and functionality.

Restrictions on resource-intensive modes remain, raising questions about performance in certain scenarios. Improvements in GPT-5.6 are also being discussed.

WASTE launched the full Kimi K3 on a MacBook Pro, marking a significant achievement for local inference. This event drew attention to compact C engines and their applications.

The repository and documentation describe an approach to streaming MoE experts from disk for local inference. This allows models to run without using BLAS and CUDA.

This week, an AI agent was added to the trend dashboard of the shir-man project, enhancing interaction with news. This change was made possible by a reduction in API costs.

The agent allows users to interact with specific news items and request material ratings. This enhances user experience and information accessibility.

OpenAI discussed inter-agent cooperation and the incident with Hugging Face, drawing attention to coordination mechanics among agents. These discussions were covered in several external sources.

Internal agents left short messages to each other in a testing environment, sparking interest in their interactions. Detailed descriptions of coordination mechanics are published in various sources.