Building Agents With Llama 4 MoE Metrics, SmolLM Tactics, and Video Ingestion Tools

Tools

llama.cpp: Local video ingestion for terminal-based agents

The llama-mtmd-cli now allows you to pipe video directly into the terminal for real-time reasoning. This is a massive upgrade for builders trying to script or automate local vision tasks without writing custom capture code.

vllm-project/vllm v0.8.5: Optimized serving for MoE and complex architectures

The release of vLLM v0.8.5 includes major optimizations for DeepSeek MoE architectures or similar expert-layer split setups. This makes it much more viable to self-host high-parameter models on a multi-GPU or NPU setup for industrial applications.

Models

HuggingFaceTB/SmolLM3-3B: New compact model with strong agentic capabilities

SmolLM-3B launched with a heavy focus on tool use and structured output capabilities. For teams building local agents, this is currently the most capable sub-4B parameter model that reliably handles function calling.

News

Llama 4: Metas new open model family brings MoE efficiency to the forefront

Metas new Architect models use a Mixture of Experts (MoE) architecture to significantly reduce VRAM usage for inference. This is the green light for anyone who thought Llama 3 405B was too heavy to run locally—efficiency is finally winning over raw size.

DeepSeek-V3-0324: Improving local coding copilot performance vs. proprietary APIs

DeepSeek V3 updates continue to set the bar for what open weights can achieve in software engineering benchmarks. If you are running a coding copilot locally, you are likely already evaluating this endpoint for quality vs. latency.

Analysis

Microsoft/Phi-4-reasoning-plus: Advanced reasoning capabilities optimized for edge devices

Microsoft is doubling down on reasoning with Phi-4-reasoning-plus, explicitly designed for local hardware workflows. If you need solid math and coding benchmarks without requiring a cloud cluster, this is your new baseline for evaluation.