1. Meta previews proprietary Muse Spark model with multimodal and reasoning capabilities
Meta has previewed Muse Spark, a proprietary multimodal model developed by its Superintelligence Labs. The model features native tool-use, visual chain of thought, and a "Contemplating mode" for parallel agentic reasoning. Benchmark results place it near frontier models like Gemini 3.1 Pro and GPT-5.4. Meta is currently offering a private API preview to select partners, with broader paid API access planned for the future. This marks a strategic shift from Meta's previous focus on open-weight Llama models.
2. Anthropic launches Claude Managed Agents in public beta for cloud-hosted agent deployment
Anthropic has released Claude Managed Agents in public beta, offering a suite of composable APIs designed to build and deploy AI agents at scale. The platform provides built-in infrastructure for secure sandboxing, authentication, state management, and tool execution. Developers can run autonomous sessions that persist through disconnections and coordinate multiple agents in parallel. The service is priced using standard Claude Platform token rates plus an active runtime fee of $0.08 per session-hour.
3. Alibaba releases Qwen3.6 Plus via API with 1M context and native vision
Alibaba has launched Qwen3.6 Plus, a proprietary multimodal model available through the Alibaba Cloud API. The model supports a 1 million token context window and native vision input, showing improved performance in agentic tasks and long-context retrieval over Qwen3.5. Pricing is set at $0.50 per million input tokens and $3.00 per million output tokens for sequences up to 256K. Alibaba has not released the model weights for self-deployment.
4. Skrun CLI tool converts agent skills into callable REST APIs
Skrun is a newly released open-source CLI tool that allows developers to deploy agent skills as callable APIs. By defining instructions in a standard SKILL.md file and configuration in an agent.yaml file, developers can expose agents via a local POST endpoint. The tool supports multiple model providers, including Anthropic, OpenAI, and Google, with automatic fallback capabilities. It also includes built-in state management and supports tool calling via local scripts or standard MCP servers.
5. Tencent releases 2B parameter Hunyuan Embodied AI vision-language model
Tencent has published the Hunyuan Embodied AI model on Hugging Face. The release is a 2-billion parameter vision-language model utilizing a Mixture-of-Transformers architecture. It is designed for embodied understanding tasks and achieves high results on benchmarks like CV-Bench and DA-2K. Developers can download and integrate the model for local vision and robotics applications.
6. Official Swift extension expands support to Cursor and other Open VSX editors
The official Swift extension is now available on the Open VSX Registry. This release enables native Swift language support, including code completion, debugging, and refactoring, in editors like Cursor, VSCodium, and AWS Kiro. Developers using agentic IDEs can now automatically install the extension without manual downloads. The update facilitates cross-platform Swift development on macOS, Linux, and Windows within a broader range of environments.
7. Memento-Skills framework enables agents to update capabilities via persistent memory
Researchers have released Memento-Skills, an open-source framework that allows large language model agents to autonomously construct and adapt task-specific capabilities. The system uses a memory-based reinforcement learning approach with stateful prompts and structured markdown files to store reusable skills. This enables agents to improve their performance through a read-write reflective learning mechanism without requiring updates to the underlying model parameters. The framework is available on GitHub for integration into custom agent workflows.
8. ThreadWeaver open-sources parallel reasoning method for 3x faster Chain-of-Thought
The ThreadWeaver project has been open-sourced, providing a new method for parallelizing reasoning tasks in large language models. The technique achieves reasoning performance comparable to standard sequential long Chain-of-Thought (CoT) models while delivering up to a 3x speedup across complex tasks. The release includes recipes for data generation, supervised fine-tuning, and reinforcement learning. Developers can use these resources to optimize inference latency for reasoning-heavy applications.
9. Upstage launches Solar Pro 3 102B MoE model via API
South Korean AI lab Upstage has released Solar Pro 3, a 102-billion parameter Mixture of Experts model with 12 billion active parameters. The model features a 128K context window and demonstrates strong performance in instruction following and agentic tool use. It is currently available as a proprietary model through Upstage's first-party API. Benchmark evaluations indicate high token usage for reasoning tasks and modest accuracy compared to frontier models.