Magic
Frontier code models with ultra-long context — LTM architecture reasoning over millions of lines.
About
Magic, founded in 2022 by Eric Steinberger and Sebastian De Ro, builds frontier-scale AI models for software engineering with a distinctive bet: ultra-long context windows. Its Long-term Memory Network (LTM) architecture powers models like LTM-2-mini, which handles 100 million tokens of context — roughly 10 million lines of code — at a fraction of the compute cost of standard attention mechanisms.
The company has raised nearly $465 million, including a $320 million round in August 2024 led by Eric Schmidt with Atlassian, Sequoia, and CapitalG participating. It partnered with Google Cloud to build two AI supercomputers (Magic-G4 and Magic-G5) targeting 160 exaflops. The stated goal is an AI coworker, not just another copilot — automating the full software engineering loop.
Worth knowing: Magic's tools have historically not been for sale — it is a research-heavy lab building toward a product. There is no public self-serve offering, so developers cannot simply sign up today. Track its model releases rather than expecting a usable tool right now.
What it can do
- LTM architecture with 100M-token context windows
- Frontier code models trained for software engineering
- Reasoning over entire codebases in a single context
- Google Cloud supercomputer partnership for training scale
- Research toward fully autonomous software engineering
Setup
No public product yet — follow research releases at magic.dev. Enterprise inquiries via the company.
Prerequisites
None for following research; enterprise engagement for any partnership.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 45; Trending score 78/100 from researcher assessment.
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.