Magic develops generative AI tools that assist software engineers with writing, reviewing, debugging and planning code, functioning like an automated pair programmer. The company uses a proprietary "Long-term Memory Network" architecture with ultra-long context windows; its LTM-2-mini model reportedly has a 100 million-token context window (about 10 million lines of code). Magic is training a larger version of that model and has partnered with Google Cloud to build two supercomputer clusters (Magic-G4 on Nvidia H100 GPUs and Magic-G5 on upcoming Blackwell chips), aiming to scale to tens of thousands of GPUs. The startup has a small team of around two dozen people and no revenue to date. Magic cites ambitions toward a path to AGI on its website and has hired senior talent including Ben Chess to expand cybersecurity, engineering, research and systems-engineering teams. Investors see a large market opportunity; Polaris Research estimates the relevant market could be worth $27.17 billion by 2032. Magic AI is a San Francisco-based startup building an AI software engineer intended to assist with complex coding tasks and act more like a coworker than a simple tool. Its technical strategy centers on handling exceptionally large context windows via a Long-term Memory Network (LTM Net) architecture; the corresponding LTM-1 model reportedly supports a 5 million context window. Operating in stealth mode, the company says it has thousands of GPUs deployed to train its next-generation models. The startup raised $117 million in a Series B and has total funding of over $145 million to date. With the new capital it is prioritizing talent recruitment and retention, seeking people who align with its stated values of integrity and innovation. The founders believe their work could help pave the way toward more expansive artificial general intelligence. Magic is building an AI-driven code-generating platform that functions like a pair programmer, able to write, review, debug and plan code changes and communicate in natural language. The product is not yet generally available; the company claims a new neural network architecture that can read 100x more lines of code than Transformers. Magic says it will flag potential license issues in generated code and will not sweep customer data into its proprietary training except for personalized systems. The startup is pre-revenue and currently has a distributed workforce of six people, with a short-term plan to grow to 25 hires focused on engineering, product and go-to-market within the next year. Magic faces competition from established tools like GitHub Copilot and broader legal and cost challenges around training large models on public code.

Total raised

$460M

Funding rounds

3

Latest round

Equity

Latest activity

Aug 2024