# Magnitude

> Run the best open models for your machine

## Overview

Magnitude is the open source inference engine for the hardware you already own. It profiles your machine, recommends the best open models for it, and tunes them for your exact hardware. One click connects your favorite agent (Pi, OpenCode, Hermes, etc.). Works on Apple Silicon, NVIDIA, AMD, or nothing but a CPU.

- **Product**: Local inference engine and desktop app for the hardware you already own
- **License**: Apache 2.0
- **Platforms**: macOS, Windows, and Linux
- **Hardware**: Apple Silicon, NVIDIA GPUs, AMD GPUs, and CPU-only machines, including unified-memory systems like DGX Spark and Strix Halo
- **Install**: Download the desktop app from https://magnitude.dev/download. The `magnitude` CLI ships inside the app; no separate install is needed.

## How It Works

1. Profiles your chip, memory, and bandwidth.
2. Estimates fit and tokens per second for every model in the catalog, then ranks models by speed, accuracy, intelligence, and memory.
3. Downloads and runs the model you choose with the built-in inference engine.
4. Tunes speculative decoding and more for your hardware.
5. Connects to your agent harness in one click.

## Features

- Models ranked by speed, accuracy, intelligence, and memory
- Hardware profiling to assess fit and estimate tokens per second before download
- Built-in inference engine; Ollama and LM Studio are not required
- Automatic speculative decoding and concurrency tuning
- Models loaded on request and unloaded when idle or memory fills
- Private, offline model execution once a model is downloaded
- No token costs, API keys, subscriptions, or rate limits
- Free and open source under Apache 2.0

## Model Catalog

The catalog at https://magnitude.dev/models lists every model Magnitude supports, with download sizes, context length, license, and intelligence score: each model's Artificial Analysis Intelligence Index score as a percentage of the top model on the index. Machine-readable copy: https://magnitude.dev/models.json. Source: https://github.com/magnitudedev/magnitude/blob/973bd9f10db3c8b37cdb530d9afc74f0635efb42/inference/catalog/models.json (updated 2026-10-14).

- [Qwen3.8 Flash Next](https://magnitude.dev/models/qwen3.8-flash-next): Qwen, 180B · 6B active, 256K context, vision, Qwen Community 1.0, from 82.6 GB. Large multimodal MoE preview of Qwen's next-generation architecture with strong agentic reasoning and low active compute.
- [Qwen3.8 27B](https://magnitude.dev/models/qwen3.8-27b): Qwen, 27B dense, 256K context, vision, Apache 2.0, from 20.5 GB. Large dense multimodal model with frontier-class agentic coding capability for its size.
- [Gemma 4 31B](https://magnitude.dev/models/gemma-4-31b-it-qat): Google, 31B dense, 256K context, vision, Apache 2.0, from 18.8 GB. Large dense Gemma model with the strongest measured coding score in its family.
- [MiniCPM-V 4.7 35B-A3B](https://magnitude.dev/models/minicpm-v-4.7-35b-a3b): OpenBMB, 35B · 3B active, 256K context, vision, Apache 2.0, from 22.5 GB. Multimodal MoE with a large knowledge footprint and low active compute for image, document, and video understanding.
- [Qwen3.6 35B-A3B](https://magnitude.dev/models/qwen3.6-35b-a3b): Qwen, 35B · 3B active, 256K context, vision, Apache 2.0, from 13.9 GB. Efficient MoE coding model with a large knowledge footprint and low active compute.
- [Muse Glimmer 30B](https://magnitude.dev/models/muse-glimmer-30b): Meta, 30B dense, 128K context, vision, Apache 2.0, from 19.6 GB. Dense multimodal model purpose-built for autonomous agentic tasks on consumer hardware.
- [Gemma 4 26B-A4B](https://magnitude.dev/models/gemma-4-26b-a4b-it-qat): Google, 25B · 3.8B active, 256K context, vision, Apache 2.0, from 15.7 GB. Mid-size MoE model balancing a substantial weight footprint with low active compute.
- [Liquid LFM2.5 24B-A2B](https://magnitude.dev/models/lfm2.5-24b-a2b): Liquid AI, 24B · 2.5B active, 128K context, LFM 1.0, from 14.5 GB. Reasoning MoE for fast on-device agents, coding, and tool use.
- [Gemma 4 12B](https://magnitude.dev/models/gemma-4-12b-it-qat): Google, 12B dense, 256K context, vision, Apache 2.0, from 7.1 GB. Mid-size dense model with native tool use, reasoning, and multimodal capability.
- [Qwen3.5 4B](https://magnitude.dev/models/qwen3.5-4b): Qwen, 4B dense, 256K context, vision, Apache 2.0, from 4.4 GB. Compact dense model for machines where responsiveness and footprint matter most.
- [MiniCPM5 2B](https://magnitude.dev/models/minicpm5-2b): OpenBMB, 2.5B dense, 128K context, Apache 2.0, from 2.2 GB. Compact dense model with native long context, tool use, and reasoning for on-device work.
- [Nemotron 3.5 Lightning 30B-A3B](https://magnitude.dev/models/nemotron-3.5-lightning-30b-a3b): NVIDIA, 30B · 3B active, 1M context, OpenMDW 1.1, from 23.6 GB. Efficient hybrid MoE model for local reasoning, coding, and agentic workflows.
- [Qwen3.5 9B](https://magnitude.dev/models/qwen3.5-9b): Qwen, 9B dense, 256K context, vision, Apache 2.0, from 8.4 GB. Small dense model with a substantial capability gain over the 4B tier.
- [Gemma 4 E4B](https://magnitude.dev/models/gemma-4-e4b-it-qat): Google, 8B dense, 128K context, vision, Apache 2.0, from 5.3 GB. Small dense model with per-layer embeddings that materially improves capability over the E2B tier while remaining practical on memory-constrained machines.
- [Liquid LFM2.5 2.6B](https://magnitude.dev/models/lfm2.5-2.6b): Liquid AI, 2.6B dense, 128K context, LFM 1.0, from 2.0 GB. Compact dense Liquid model tuned for fast on-device tool use and coding workflows.
- [MiniCPM5 1B](https://magnitude.dev/models/minicpm5-1b): OpenBMB, 1.1B dense, 128K context, Apache 2.0, from 0.7 GB. Compact dense model for on-device reasoning and tool use.
- [Gemma 4 E2B](https://magnitude.dev/models/gemma-4-e2b-it-qat): Google, 5.1B dense, 128K context, vision, Apache 2.0, from 3.7 GB. Very small dense model with per-layer embeddings optimized for on-device use.
- [Liquid LFM2.5 8B-A1B](https://magnitude.dev/models/lfm2.5-8b-a1b): Liquid AI, 8.3B · 1.5B active, 125K context, LFM 1.0, from 5.5 GB. Low-active-parameter Liquid MoE optimized for fast local reasoning and tool use.
- [MiniCPM-V 4.7 1B](https://magnitude.dev/models/minicpm-v-4.7-1b): OpenBMB, 1B dense, 256K context, vision, Apache 2.0, from 1.8 GB. Compact multimodal model for on-device image, document, and video understanding.

## Supported Agent Harnesses

Pi, OpenCode, Hermes, OpenClaw, Codex, Claude Code, Oh My Pi, and Cline. Pick a model and connect your harness in one click.

## FAQ

- **What is Magnitude?** An open source inference engine optimized for consumer hardware. The desktop app profiles your machine, recommends the best models for it, then downloads, tunes, and runs them. One click connects the agent you already use.
- **How does it know what my machine can run?** Magnitude profiles your hardware and estimates tokens per second for every model in the catalog before you download anything. It ranks them by speed, accuracy, intelligence, and memory so you can pick.
- **How is this different from Ollama or LM Studio?** They run whatever model you pick. Magnitude helps you pick. It estimates how every model and quant will perform on your machine before you download, then tunes the one you choose for your exact hardware, from context size to speculative decoding.
- **What hardware do I need?** There's no fixed minimum. Magnitude profiles your machine and recommends what runs well on it. More memory lets you run larger models.
- **What systems does Magnitude support?** The desktop app is native on macOS, Linux, and Windows. It runs on Apple Silicon, NVIDIA and AMD GPUs, and CPU-only machines.
- **Do I need to manage it after setup?** No. It runs in the background, loads models when your agent needs them, and unloads them when idle or memory gets tight.
- **Is it private?** Yes. Prompts, files, and models stay on your machine. Once a model is downloaded, no internet connection is needed.

## Links

- Website: https://magnitude.dev
- Download: https://magnitude.dev/download
- Model catalog: https://magnitude.dev/models
- Documentation: https://docs.magnitude.dev
- GitHub: https://github.com/magnitudedev/magnitude
- Twitter/X: https://x.com/usemagnitude
- Contact: founders@magnitude.dev