Llama 4 Drops, Phi-4 Thinks Locally, and Inference Tools Get a Boost

Tools

Microsoft's Phi-4 Reasoning Plus: 14B Model with Deep Thinking for Local Use

Phi-4 Reasoning Plus is Microsoft's new 14B model that actually thinks step by step for tougher stuff like math and code. It's small enough to run locally on decent hardware, which is a serious move for builders who want reasoning power without calling an API.

vLLM v0.8.5: Faster MoE Kernels, Llama 4 Support, TPU Upgrades

vLLM v0.8.5 adds Llama 4 and Phi-4 MoE support, plus significant speed boosts for MoE kernels and structured output on TPUs. If you're self-hosting any of these models, this is your must-upgrade release.

SmolLM3-3B: A Tiny Powerhouse for Local, Tool-Calling AI

SmolLM3-3B is a compact model with surprisingly strong tool-calling and instruction-following for its size. It's optimized for edge devices and local inference, making it a practical choice for builders embedding AI in apps without cloud dependency.

News

Meta Open-Sources Llama Guard 4 for Multimodal Content Safety

Meta just open-sourced Llama Guard 4, a multimodal safety model that judges both text and images. For builders running local pipelines, this is a ready-made safety layer you can integrate now instead of building from scratch.

Llama 4 Model Family Launches: Four Sizes, MoE Architecture, Open Weights

Meta dropped Llama 4 in four sizes (Scout, Maverick, plus two more coming). Scout is 109B total but uses 17B active parameters via MoE, and you can run it on a single H100. Maverick is 400B but beats GPT-4o on many benchmarks. This is the most serious open-weights competition to frontier models yet.

DeepSeek-V3-0324: New Open Model Challenges GPT-4o on Reasoning and Code

DeepSeek-V3-0324 continues the trend of Chinese labs shipping competitive open models. This version focuses on better reasoning and coding, directly benchmarking against GPT-4o. Builders running local inference have another serious model to test.