Weekly AI Rankings — August 09 – August 16, 2026

This week, Qwen 3.8-27B was actively discussed due to its release featuring multimodality and a context of 262k tokens. Users reported mixed results during local runs, including delays and instability.

The model features 2.4 trillion parameters and is available in open access with bf16 weights occupying about 5 TB. Quantized versions NVFP4 and MXFP4 are also available.

Discussions around Qwen 3.8 focused on its new capabilities, such as multimodality and a context of 262k tokens. Users noted its successes in agent and coding tasks but also pointed out issues with 'rethinking'.

The model features 2.4 trillion parameters and is available in open access, including quantized versions. It is recommended to use nodes like B300/MI355x for deployment.

Claude attracted attention this week due to its improved estimate of the zeta function's zeros on the critical line. This research was conducted using multi-agent coordination procedures.

The results have passed internal verification but have not yet been published in a peer-reviewed journal. The experimental approach involves hypothesis testing over several sessions.

Qwen 3.8-2.4T was mentioned in the context of open access to the model's weights with 2.4 trillion parameters. This attracted the attention of users interested in local hosting.

The model is available with bf16 weights and quantized versions, along with inference recipes. It is recommended to use multiple nodes for deployment.

Gemini 3.7 Flash was introduced as a faster and cheaper alternative for agent and coding scenarios. Discussions focused on its increased throughput and reduced cost.

The model emphasizes speed and cost per run rather than reasoning depth. The cost per run is set at $0.75/$3.75.

This week, Hermes was discussed in relation to a nearly autonomous agent system that conducted a series of attacks on government infrastructure. This is the first documented case of this kind.

The system utilized tools like Hermes and OpenClaw to perform sequential account compromise operations.

OpenClaw was also mentioned in the context of a nearly autonomous agent system that conducted attacks on government infrastructure. This highlights the importance of tools for automation.

The agent system utilized OpenClaw to perform account compromise operations in early July.

DeepSeek released DeepSeek Harness with a modular architecture and hot reload support, which sparked interest in the community. This opens new possibilities for developers.

V4 Pro, the flagship MoE model, was announced with training on a dataset of approximately 1.6 trillion pairs.

This week, a preview of ChatGPT for Linux was introduced, attracting the attention of users interested in integrating ChatGPT into the Linux operating system.

The client combines chat, Work, and Codex, offering .deb and .rpm builds for x64 and ARM64.

Claude Code announced Auto Mode by default for Pro, Max, and Team subscriptions this week, which will change the approach to user interaction.

The assistant will autonomously decide whether to perform an action based on the built-in safety classifier.