Open Model Anatomy & The Agentic SmolLM

Tools

llama.cpp Adds Native Llama 4 and DeepSeek-V3 MoE Support

The latest llama.cpp update is a huge deal for local builders. It lets you run large MoE models, like Llama 4 Scout, on much less RAM, which is a direct path to testing frontier models on consumer hardware.

Microsoft's Phi-4 Reasoning Model is Now Open on Hugging Face

Don't sleep on Microsoft's Phi-4-reasoning-plus model. It's a smaller model fine-tuned for complex reasoning tasks, designed to run locally and give you a taste of 'thinking' capabilities without needing a massive cloud deployment.

News

SmolLM3: Hugging Face Drops a 3B Parameter Model with Strong Agentic Capabilities

SmolLM3 is a new small language model that punches above its weight for tool calling. It's optimized for real-world agentic tasks, making it a practical choice for developers building AI systems that need to interact with other tools and APIs.

Meta Releases Llama 4: A New Family of Open, Efficient MoE Models

Meta dropped Llama 4, their latest open model family with a Mixture-of-Experts architecture. The 'Scout' version is built for efficiency and runs locally on a single GPU, making it a strong contender for on-device and edge AI applications.

Analysis

New Research Explores Cost-Effective Mixture of Agents Architecture

This paper proposes using a smaller, smarter model to decide when to invoke a larger one, optimizing cost and latency. It's a foundational tactic for building practical, efficient AI systems that don't spend a fortune on inference.