Local Deployment Guide: Agentic Models (Muse Glimmer & LFM2.5-2.6B)

This guide covers requirements and optimization strategies for running high-performance agentic models locally on PC, Mac, and Mobile hardware.

Hardware Requirements

• For Muse Glimmer (30B): Requires a device with 24 GB to 32 GB of memory (e.g., single consumer GPU or Apple Silicon M-series).
• For LFM2.5-2.6B (Lightweight Agenting): Low resource footprint (

Installation & Launch Guide

1. Environment Setup: Ensure your system supports compatible backends including llama.cpp, MLX, ONNX, vLLM, SGLang, hoặc ExecuTorch depending on target architecture.
2. Model Weight Acquisition: Download weights from Hugging Face which include open access formats like Apache 2.0 licensed weightings or ODS automated setup scripts via GitHub/Osmantic/ODS.
3. Deployment Execution (via CLI tools such as Ollama or LM Studio):

   # Example pattern for local inference using standard engines
   $ [engine_name] run [model_id]

   Please note that specialized services like ODS can automatically detect PC characteristics and select appropriate models.

Optimization & Performance Tips

• Hardware Acceleration: Use DFlash optimization if available certain hardware may see speedups up to 3.1x (RTX 5090) or 1.8x (M5 Max).
• Throughput Benchmarks / Optimization targets: aim for high-speed generation settings.
  - M5 Max: ~220 tokens per second (LFM)
  - Ryzen AI Max+ 395: ~113 tokens per second (LFM)
  - Smartphone level: Maintain ~30 tokens per second (LFM)
• Specialized Engines: For maximum throughput on NVIDIA H100, utilize vLLM/vllm architectures capable of reaching higher token thresholds.

Bottom Line: Running these agentic pipelines locally ensures full privacy with no cloud subscription fees while enabling complex tool calling and long chain actions directly on your own machine.

! DYOR (Do Your Own Research)