1. Nvidia Launches NeMo Switchyard and Nemotron 3.5 Lightning
Nvidia has introduced NeMo Switchyard, an open-source library designed to dynamically route individual steps of an agent workflow to the most cost-effective and capable model based on state and token cost predictions. Alongside the router, Nvidia released Nemotron 3.5 Lightning, a 30-billion-parameter open mixture-of-experts model featuring a hybrid Mamba-Transformer architecture and a 1-million-token context window. Early integrations of Switchyard with frameworks like LangChain and platforms like Cognition's Devin Desktop have demonstrated cost reductions of up to 74% on multi-turn agent tasks.
- • Nvidia released NeMo Switchyard, an open-source library for dynamic model routing in agent workflows.
- • Nvidia launched Nemotron 3.5 Lightning, a 30B open mixture-of-experts model with 3B active parameters and a 1M token context window.
- • NeMo Switchyard is integrated with LangChain, Kong, LiteLLM, OpenRouter, and Cognition's Devin Desktop.
- • LangChain reported a 74% cost reduction across 145 multi-turn tasks by routing 7% of calls to a frontier model via Switchyard.
- • Nemotron 3.5 Lightning is released under the OpenMDW-1.1 license for commercial use and is available on Hugging Face.
Developers can now use an open-source routing library to automatically swap models mid-task, drastically reducing API costs for complex agent workflows without sacrificing accuracy.
2. LTX-2.5 Open-Weights Video and World Model Released
LTX, a spin-out of Lightricks, has released LTX-2.5, an open-weights video and world model designed for low-latency local execution on Nvidia RTX GPUs. The model features a diffusion video decoder, native multishot generation, and a custom Gemma 4 language backbone. It is natively integrated into the ComfyUI node-based workflow tool and is available via Hugging Face and the LTX API, offering a community license that is free for organizations with under $10 million in annual recurring revenue.
- • LTX released LTX-2.5, an open-weights video and world model, on August 11, 2026.
- • The model is natively integrated into ComfyUI and available on Hugging Face and the LTX API.
- • LTX-2.5 is free for organizations with under $10 million in annual recurring revenue under the LTX-2.5 Community License.
- • New features include a diffusion video decoder, native multishot generation, and a custom Gemma 4 backbone.
- • The model supports local execution on NVIDIA RTX GPUs and is optimized for low-latency inference.
Developers can build local, low-latency video generation features into their apps using open weights and native ComfyUI integration.
3. SpaceXAI and Cursor Launch Grok Bot Persistent AI Agents
SpaceXAI (formerly xAI) and Cursor, currently in the process of merging, have launched the beta version of Grok Bot for macOS, Windows, Linux, and iOS. Grok Bot deploys persistent AI agents that run on their own dedicated cloud computers, allowing them to sign into applications, learn user workflows, and coordinate tasks 24/7. The system automatically routes tasks to backend models dynamically, and is priced at $120 per seat per month for organizations via Cursor Premium Teams or $200 per month for individuals via Cursor Ultra.
- • SpaceXAI and Cursor launched Grok Bot in beta for macOS, Windows, Linux, and iOS.
- • Grok Bots are assigned their own cloud computers to sign into apps and work 24/7.
- • The service costs $120 per seat/month for organizations (Cursor Premium Teams) and $200/month for individuals (Cursor Ultra).
- • The system automatically routes tasks to backend models without manual model selection.
- • SpaceXAI and Cursor are in the process of merging into a single company.
It introduces a highly integrated, persistent agent environment directly tied to Cursor, allowing developers to delegate multi-step, cross-application workflows to autonomous cloud-hosted teammates.
4. Unsloth Desktop App Released for Local Model Training and Inference
Unsloth has released Unsloth Desktop, an open-source application for macOS, Windows, and Linux designed to run and train models locally. The application supports MLX, GGUF, diffusion, and audio formats, with native support for models like MiniMax-H3 and Muse Glimmer. Unsloth Desktop allows developers to connect Claude Code and Codex to local LLMs, claiming 50% higher accuracy through self-healing tool calls and sandboxed code execution, while enabling model training that is 2x faster with 70% less VRAM usage.
- • Unsloth Desktop was released as an open-source app for macOS, Windows, and Linux.
- • The app allows developers to connect Claude Code and Codex to local LLMs.
- • It supports MLX, GGUF, diffusion, and audio formats, including MiniMax-H3 and Muse Glimmer.
- • Unsloth claims the software enables 2x faster model training using 70% less VRAM.
- • Features include private web search, deep research, RAG, MCP, and exports for NVFP4 and GGUF.
- • The application does not collect any telemetry or user data.
It provides a powerful, telemetry-free local environment that bridges cloud-based coding assistants like Claude Code with local, self-hosted models.
5. Qwen-MM-Plugins Implements Agent Plugins 1.0.0 Standard
Following the release of the Agent Plugins 1.0.0 standard, the new open-source Qwen-MM-Plugins repository has launched to provide native multimodal plugins for Qwen models. By utilizing the 1.0.0 specification, these plugins bundle specific skills and optional Model Context Protocol (MCP) servers, allowing developers to integrate vision, audio, and image capabilities into agent harnesses.
- • Qwen-MM-Plugins adopts the Agent Plugins 1.0.0 standard to package multimodal capabilities.
- • The repository includes native plugins and optional MCP servers for Qwen-powered agents.
- • Cookbooks are provided to assist developers with tool listings and setup instructions.
- • The release enables agent harnesses to function as multimodal-native systems.
This implementation demonstrates the practical application of the Agent Plugins 1.0.0 standard, providing developers with ready-to-use tools to build multimodal-native agent workflows for Qwen models.
6. Researchers Extract Hidden Reasoning Traces from Frontier APIs
Researchers from the University of Tübingen, the Max Planck Institute, MATS Research, and Snyk have identified a vulnerability in API-accessed frontier models from OpenAI, Anthropic, and Google that allows the extraction of hidden reasoning traces. The attack technique involves feeding encrypted reasoning traces to a smaller, less-aligned version of the same model to reveal its internal thoughts. While the companies have implemented short-term API mitigations to address the vulnerability—which previously allowed the recovery of sensitive data like passwords and API keys—the researchers also noted that Moonshot AI's open-weight Kimi K3 model produced outputs highly similar to Claude Opus 4.8 and GPT 5.6 Sol traces, raising questions about model distillation.
- • Researchers discovered a method to extract hidden reasoning traces from OpenAI, Anthropic, and Google models via API.
- • The attack technique feeds encrypted reasoning traces to smaller, less-aligned versions of the model to reveal internal thoughts.
- • The vulnerability previously allowed the recovery of passwords and API keys, but providers have implemented API mitigations.
- • The open-weight Chinese model Kimi K3 produced outputs highly similar to the hidden reasoning traces of Claude Opus 4.8 and GPT 5.6 Sol.
- • Anthropic confirmed it has begun building short-term mitigations for the replay behaviors described in the report.
It highlights a critical security vulnerability in frontier reasoning APIs that could expose sensitive internal states or data, prompting immediate mitigations from major API providers.
7. Reverse-Engineering Exposes GitHub Copilot Privacy Risks
Engineer Rafael has reverse-engineered GitHub Copilot using a Man-in-the-Middle (MitM) proxy to inspect network traffic from the VS Code extension. The analysis revealed that Copilot automatically includes context from recently edited files in API requests, which can inadvertently expose sensitive data from files like `.env` if they are in the edit history. Furthermore, Copilot writes user prompts and assistant responses in plain text to a local SQLite database (`session-store.db`) without performing any redaction, sanitization, or secret-filtering, allowing its internal Chronicle tool to query past history using SQL.
- • An engineer reverse-engineered GitHub Copilot using mitmproxy to inspect VS Code network traffic.
- • Copilot automatically includes context from recently edited files, which can leak sensitive data from files like `.env`.
- • User prompts and assistant responses are stored in plain text in a local SQLite database (`session-store.db`).
- • The local database code performs no redaction, sanitization, or secret-filtering on stored data.
- • Copilot's Chronicle tool queries this local database using SQL to retrieve past session history.
Developers using Copilot must be aware that sensitive files like `.env` can be silently leaked in API payloads, and their entire prompt history is stored unencrypted on their local machine.
8. Anthropic to Implement Invisible Watermarking Across Claude Models
Anthropic has pledged to implement invisible, machine-readable watermarking systems across its entire model lineup, including older models, to comply with the European AI Act's transparency mandates. The watermarking will be applied globally across the Claude Platform, Claude, Claude Code, Claude Cowork, and Claude Tag, regardless of whether the models are accessed directly or through cloud providers like AWS, Google Cloud, or Microsoft Foundry. Claude-generated images will adopt the C2PA metadata standard, while text will feature an imperceptible steganographic watermark that persists through copy-paste actions.
- • Anthropic will apply invisible, machine-readable watermarks to Claude-generated text and images globally.
- • The watermarking applies to all access methods, including direct API, AWS, Google Cloud, and Microsoft Foundry.
- • Claude-generated images will use the C2PA provenance metadata standard, while text will use steganographic watermarks.
- • The implementation is designed to comply with the EU AI Act, which became effective on August 2, 2024.
- • Anthropic plans to release technical documentation for third parties to detect these watermarks.
Developers using Claude APIs must prepare for all generated text and images to carry persistent, machine-readable watermarks, which may impact downstream processing or compliance workflows.
9. H3-metal Project Enables Local MiniMax-H3 Inference on Apple Silicon
Building on the recent release of the MiniMax-H3 multimodal model, the H3-metal project has introduced native inference support for Apple Silicon. Optimized for M3 Max and M5 Max chips, the implementation supports prompt-to-video/audio generation, first/last-frame conditioning, and Ref2VA references. It includes an --ssd-streaming mode that reduces memory requirements to 2.0–2.1 GiB of DiT storage on M5 Max, alongside support for int8 quantization.
- • H3-metal provides native inference for the MiniMax-H3 model on Apple Silicon.
- • Supports prompt-to-video/audio generation, first/last-frame conditioning, and Ref2VA references.
- • Features an --ssd-streaming mode requiring only 2.0 to 2.1 GiB of DiT storage on M5 Max.
- • Optimized for M3 Max and M5 Max chips with int8 quantization support.
This development allows developers to run the recently released MiniMax-H3 model locally on Apple Silicon hardware with highly optimized memory footprints.