Magnitude · Launch Video Breakdown: Hook, Pacing & Motion Design

Scene-by-scene timeline & spoken transcript

1. 0:00 – 0:04

   The Hook

   Local Model Selection & Download

   “(No spoken dialogue — silent video with ambient visual sound design)”

   On screen

   Magnitude The best way to code with local models Current directory: ~/magnitude Tip: Use /settings to manage models. Choose a local model Apple M4 Max · macOS ARM64 · 16 cores · 64 GiB unified · MTL AVAILABLE TO DOWNLOAD > Qwen3.6 35B-A3B Balanced Qwen3.6 27B Best Quality Gemma 4 26B-A4B Fastest Gemma 4 12B Lightweight Q6_K · 32.6 GB · 200K ctx ~37 tok/s at 200K ctx Best overall mix of coding ability, speed, and memory use. Runs at about 36.9 tokens/ sec with 200K context and retains very high quality with minimal loss. You can switch models or download more anytime from /settings. Select a model to start coding... No model configured Downloading Qwen3.6 35B-A3B · Q6_K 19% 6.33 GB / 32.6 GB 24 MB/s · about 18 minutes remaining Cancel (Esc) Download in progress · Esc cancel Downloading Qwen3.6 35B-A3B... Qwen3.6 35B-A3B High / 200k ON THIS COMPUTER > Qwen3.6 35B-A3B Load Loading Qwen3.6 35B-A3B into memory : Loading model weights... Loading model into memory... : Loading model · 1% · Stop Loading Qwen3.6 35B-A3B... : Loading model · 92% · Stop

   Camera

   Static, full-screen terminal window view, with subtle UI element interactions.

   Motion

   Typing animation, progress bar fill, cursor interaction, and text highlight.

2. 0:04 – 0:06

   Problem Agitation

   Offline Mode Demonstration

   “(No spoken dialogue — silent video with ambient visual sound design)”

   On screen

   Wi-Fi Set up Wi-Fi to wirelessly connect your Mac to the internet. Turn on Wi-Fi, then choose a network to join. Learn More... Wi-Fi is off Ask to join networks Known networks will be joined automatically. If no known networks are available, you will have to manually select a network. Ask to join hotspots Allow this Mac to automatically discover nearby personal hotspots when no Wi-Fi network is available. Advanced... Chat with the agent... Qwen3.6 35B-A3B High / 200k 40.7 GB mem Explore this repo and tell me what it does

   Camera

   Zoom-in on a macOS settings panel, then a quick cut back to the terminal.

   Motion

   Mouse cursor interaction, toggle switch animation, and rapid scene cut.

3. 0:06 – 0:12

   Product Reveal

   AI Agent Repo Exploration

   “(No spoken dialogue — silent video with ambient visual sound design)”

   On screen

   Working · 0:03 · Prefill · 3.4k / 6.2k tokens Type to queue a message... Qwen3.6 35B-A3B High 5k / 200k (3%) 40.7 GB mem Magnitude The best way to code with local models Current directory: ~/magnitude Tip: Use /settings to manage models. Explore this repo and tell me what... Explore this repo and tell me what it does Working · 0:06 · Thinking Type to queue a message... Qwen3.6 35B-A3B High 5k / 200k (3%) 40.7 GB mem Let me explore the repo structure and key files to give you a thorough overview. → Read 2 files (README.md, package.json) Working · 0:08 · Prefill · 512 / 2.6k tokens · 6.2k cached Type to queue a message... Qwen3.6 35B-A3B High 9k / 200k (5%) 40.7 GB mem List files (packages) Working · 0:11 · Prefill · 0 / 216 tokens · 8.8k cached

   Camera

   Static terminal view, focusing on text input and AI agent response.

   Motion

   Typing animation, text reveal, progress indicators, and subtle cursor movement.

4. 0:12 – 0:35.8

   Feature Teaser

   Deep Dive into Magnitude's Architecture

   “(No spoken dialogue — silent video with ambient visual sound design)”

   On screen

   Explore this repo and tell me what... Let me explore the repo structure and key files to give you a thorough overview. → Read 2 files (README.md, package.json) List files (packages) Let me get a cleaner view of the packages and read key architectural files. $ ls packages/ | head -60 ✓ acn acn-dashboard acn-protocol ... 26 lines collapsed utils vcs version $ for pkg in packages/*/package.json; do echo "=== $pkg ==="; cat "$pkg" | grep... Working · 0:13 Type to queue a message... Qwen3.6 35B-A3B High 9k / 200k (5%) 40.7 GB mem $ for pkg in packages/*/package.json; do echo "=== $pkg ==="; cat "$pkg" | grep... === packages/acn-dashboard/package.json === "name": "@magnitudedev/acn-dashboard", "version": "0.0.1", ... 82 lines collapsed === packages/version/package.json === "name": "@magnitudedev/version", "version": "0.0.1", Here's what I found: Magnitude – Open Source Coding Agent with Local Inference Magnitude is an AI coding agent platform with its own custom inference engine built in Rust. It runs locally on your machine (macOS/Linux, WSL on Windows) and can also connect to cloud-hosted models via Magnitude Cloud. Core Architecture CLI/Web (clients) → client-common → SDK → ACN (daemon) The system is split into three major layers: 1. Inference Engine (Rust – inference/ ) A custom inference engine built on llama.cpp, handling: - Model profiling and hardware detection - Memory calculation before loading - Acceleration, placement, and batching tuned for your hardware - Parallel agents that retain full context windows - Native prefill, cache reuse, and generation performance metrics The engine is organized as Rust crates under inference/crates/ (e.g., icn-engine , icn-models , icn-server ). 2. Agent Runtime ( packages/agent/ ) The coding agent itself – it can: - Inspect and edit files - Run shell commands - Work with images - Manage long-running tasks - Delegate work to parallel agents - Switch models while preserving consistent tool use Built around event sourcing ( event-core ) with named projections (Window, Display, TaskGraph, etc.) for replaying and inspecting session state. 3. Client Layer ( cli/ , web/ , packages/client-common/ ) - CLI: TUI built with OpenTUI (React-based terminal UI) - Web: Web-based client - client-common: Shared state, hooks, display sync between client and daemon Package Ecosystem (30+ packages) Layer Key Packages Purpose Client cli, web Terminal and web UIs Shared client-common, sdk, logger, utils State hooks, RPC client, tracing Daemon acn, acn-protocol Local daemon server + RPC wire contract Agent agent, roles Coding agent runtime + worker specialization AI ai, providers Provider-agnostic model contracts + concrete providers Inference icn, icn-protocol Inference lifecycle wrapper + protocol Event event-core, tracing Event sourcing, projections, OTEL tracing System storage, version Persistent session/config/auth, version generation Tooling skills, vcs, ripgrep, shell-classifier Agent tools and utilities Release release Build/release pipeline Key Technical Details - Language: TypeScript/React (clients + most packages) + Rust (inference engine) - Runtime: Bun (package manager + test runner) - Effect-TS: The entire codebase uses Effect-TS for async and dependency injection - RPC: Custom typed RPC system via acn-protocol (gRPC-like, over HTTP) - Event Sourcing: All agent sessions are event-sourced with replayable projections - Telemetry: OpenTelemetry via local collector "Motel" for tracing/spans/logs - License: Apache 2.0 - Package Manager: Bun workspaces with Turborepo for builds In short: it's a full-stack AI coding agent platform where the inference is done locally (with cloud fallback), the agent is event-sourced, everything is Effect-TS, and the architecture cleanly separates the daemon (ACN), the agent runtime, and the client layers. Qwen3.6 35B-A3B worked for 17 seconds · 67.8 tok/s Chat with the agent...

   Camera

   Static terminal view, with continuous scrolling and text generation.

   Motion

   Typing animation, text reveal, code block expansion, and smooth vertical scrolling.