LLMs

OpenAI has released the Decisions API into public beta, offering a new interface for structured reasoning and decision-making workflows. The announcement, highlighted by significant engagement on Hacker News, marks a shift towards more deterministic outputs from their models. Developers can now access this feature directly through the official developer documentation.

• Decisions API is now available in public beta for broader developer access.
• The release signals a focus on structured reasoning capabilities in LLMs.
• High community engagement suggests strong interest in this new feature set.
• Integration details are documented in the official OpenAI developer guides.

DeepSeek has open-sourced DeepGEMM, a unified high-performance tensor core kernel library designed for modern large language models. It consolidates primitives like FP8/FP4/BF16 GEMMs, fused MoE with overlapped communication, and MQA scoring into a single CUDA codebase. The project uses DeepJIT for runtime compilation, eliminating the need for CUDA compilation during installation, and offers a simplified alternative to CUTLASS for learning and implementation.

• Supports FP8, FP4, and BF16 GEMMs plus fused MoE with Mega MoE communication overlap.
• Uses DeepJIT for runtime compilation, removing build-time CUDA dependencies.
• Simplifies kernel development by avoiding heavy CUTLASS/CuTe template complexity.
• Includes specialized primitives like MQA scoring for lightning indexing.

Strands has released Decider 2B, a small, open-source model designed specifically for decision-making tasks within agent workflows. The release positions this 2-billion parameter model as a lightweight alternative for routing and logic execution in AI agents. It is part of the broader Strands agents ecosystem aimed at simplifying autonomous agent development.

• Decider 2B is an open-source model focused on decision logic for agents.
• At 2B parameters, it offers a lightweight option for routing tasks.
• Released by Strands as part of their agents ecosystem.
• Targets practitioners building autonomous agent workflows.

Google has introduced EmbeddingGemma 2, an open-source model designed for generating embeddings from multimodal data. This release provides developers with a lightweight alternative for processing diverse data types within their AI pipelines. The model aims to simplify the integration of multimodal capabilities into existing applications.

• Google releases EmbeddingGemma 2 as an open, lightweight multimodal embedding model.
• The tool enables developers to process diverse data types in AI pipelines.
• Provides a lighter alternative for integrating multimodal capabilities into apps.