xAI Integrates Grok with Nous Hermes Agent, Opens X Search and Premium Access

Available in Hermes Agent today: X Premium subscription support and the ability to search X platform posts directly from the agent interface.

xAI announced a major integration connecting Grok to Nous Research's open-source self-improving Hermes Agent. Users can now use their X Premium subscription within the agent and search X posts natively. The move bridges xAI's flagship model with the broader agent ecosystem, marking a step toward more deeply interconnected AI tooling across platforms.

vLLM v0.21.0 Ships with 367 Commits Across 202 Contributors

vLLM v0.21.0 introduces KV Offload with HMA, a speculative decoding thinking budget for reasoning models, TOKENSPEED_MLA on Blackwell GPU for DeepSeek R1 and Kimi K2.5, Mooncake distributed KV, and DeepSeek V4 pipeline parallelism. The release also moves the codebase to C++20. With 49 new contributors joining this cycle, the project continues to accelerate as the inference backbone for open-source LLMs.

30B-A3B Reasoning Model Reaches Gold Medal Level Across Physics and Math Olympiads

Co-first authors have released a 30B-parameter reasoning model with 3B active parameters that achieves gold medal level results on both physics and mathematics Olympiad evaluations. The compact architecture demonstrates that reasoning capability does not require massive parameter counts, challenging the scale-above-all assumption in current model development.

Raschka Publishes Visual Tour of LLM Architecture Advances: KV Sharing to Compressed Attention

Sebastian Raschka published a comprehensive blog article analyzing recent LLM architecture innovations for long-context efficiency. Covering Gemma 4's cross-layer KV sharing and per-layer embeddings, Laguna XS.2's layer-wise attention budget allocation, ZAYA1-8B's compressed convolutional attention, and DeepSeek V4's multi-head compression, the piece maps the emerging design patterns that reduce KV cache size and memory bandwidth to enable longer reasoning contexts.

Codex Keyboard Shortcuts Now Customizable

OpenAI updated Codex based on user feedback, allowing users to customize keyboard shortcuts from settings to match their personal workflow instead of adapting to defaults.

DCI Retrieval Method Tops Hugging Face Daily Papers

The Direct Corpus Interaction method enables agents to search raw text using common tools like grep and bash, eliminating the need for embedding models, vector indexes, or offline indexing. Tests on BRIGHT, BEIR, BrowseComp-Plus and multi-hop QA benchmarks showed average gains of 11% to 30.7% over traditional sparse, dense, and reranking baselines. The findings suggest retrieval quality depends as much on the resolution of the model-to-corpus interface as on reasoning ability.

Replit AI Agent Helps Designer Ship Nearly Daily for 18 Months

A designer with no coding background has been using Replit's AI Agent to ship products almost daily for a year and a half, showcasing how low-code AI tools are lowering the barrier from idea to production deployment.