Magic.dev
** - (USA) A startup building an AI "coworker" for software engineers, designed to handle complex coding and engineering tasks.
Last verified:
What is Magic.dev?
Magic.dev is an AI research company building frontier code models to automate software engineering and AI research, with the ultimate goal of developing safe Artificial General Intelligence (AGI). The company's core product is a advanced AI coding system that goes beyond traditional code assistants by acting as a full coworker rather than just a copilot, capable of planning, writing, reviewing, and testing code across entire codebases.
The platform's standout feature is its ultra-long context window of 100 million tokens (LTM-2-Mini model), which equals approximately 10 million lines of code or 750 novels. This allows the AI to reason over entire repositories, documentation, and tickets in a single pass. The system combines frontier-scale pre-training, domain-specific reinforcement learning, ultra-long context processing, and inference-time compute to produce reliable code contributions that survive CI and code review.
Magic.dev is designed for software engineering teams, AI research organizations, and enterprises working on large-scale codebases that need end-to-end automation. The company operates its own supercomputing infrastructure with thousands of NVIDIA GB200 GPUs and partners with Google Cloud for scaling. Access is through direct enterprise engagement rather than self-serve product offerings.
Magic.dev pricing
Pricing model: Freemium
Contact for pricing. Magic.dev is a frontier AI research company focused on automating software engineering and AI research. No public pricing tiers, free plans, or self-serve product offerings are listed on the website. The company operates on a contact/enterprise basis with direct engagement from sales. The company has raised $515 million total including $320M recent investment from Eric Schmidt, Jane Street, Sequoia, Atlassian, Nat Friedman, Daniel Gross, Elad Gil, and CapitalG.
Magic.dev pros
- 100 million token context window processes entire codebases in one pass
- LTM mechanism is 1000x cheaper than Llama 3.1 405B attention for long context
- Requires fraction of one H100 vs 638 H100s per user for same context
- Domain-specific reinforcement learning tuned for code generation
- Frontier-scale pre-training for advanced software engineering automation
- Inference-time compute enhances reasoning and output quality
- Automates entire software engineering workflow end-to-end
- Can plan, write, review, and test code across large codebases
- Reasons over repos, docs, and tickets together for context
- Produces reproducible evaluations with safer pilots
- Clearer audit trails for code contributions
- Partnership with Google Cloud for GB200 NVL72 cluster
- Operates 8000+ H100s and thousands of GB200 GPUs
- $515M raised from top investors including Sequoia, CapitalG, Nat Friedman
- AGI Readiness Policy for evaluating and reducing existential AI risks
- HashHop evaluation benchmark for authentic long-context testing
- Built by small team of 23 engineers and researchers focused on AGI
Magic.dev cons
- No public pricing tiers listed - contact for pricing only
- No free tier or self-serve product offering available
- No public API or consumer-facing web interface published
- Access limited to direct enterprise engagement
- Primarily aimed at organizations not individual developers
- Small prototype model code synthesis abilities not good enough yet
- Currently 23 people - limited team size for development
- No IDE integration mentioned like traditional coding assistants
- Still in research phase rather than mature commercial product
Frequently asked questions about Magic.dev
What is Magic.dev?
Magic.dev is an AI company building frontier code models to automate software engineering and research. The company is working toward building safe AGI by automating AI research and code generation to improve models and solve alignment more reliably than humans can alone.
How large is Magic's context window?
Magic supports 100 million token context windows. LTM-2-Mini is their first model with this capability, which equals approximately 10 million lines of code or 750 novels. This is about 1000x greater than frontier models from other providers.
What is Magic's approach to achieving AGI?
Magic aims to automate AI research and code generation using frontier code models. They combine frontier-scale pre-training, domain-specific reinforcement learning, ultra-long context up to 100M tokens, and inference-time compute to improve models and solve alignment safely.
What is Magic's AGI safety approach?
Magic believes the safest path to AGI is automating AI research and code generation to improve models and solve alignment more reliably than humans. They maintain an AGI Readiness Policy to evaluate, monitor, and reduce existential risks of AI capabilities now and in the future.
Which cloud provider does Magic partner with?
Magic partners with Google Cloud to build their next-generation AI supercomputers. They are building Magic-G4 powered by NVIDIA H100 Tensor Core GPUs and Magic-G5 powered by NVIDIA GB200 NVL72 on Google Cloud, with ability to scale to tens of thousands of Blackwell GPUs.
What is the LTM (Long-Term Memory) mechanism?
LTM is Magic's proprietary architecture for ultra-long context processing. For each decoded token, LTM-2-mini's sequence-dimension algorithm is roughly 1000x cheaper than Llama 3.1 405B's attention mechanism at 100M context. LTM requires a small fraction of one H100's HBM per user versus 638 H100s per user for Llama 3.1 405B.
What is HashHop?
HashHop is a new evaluation benchmark designed by Magic to eliminate implicit and explicit semantic hints in long-context testing. It uses incompressible random hashes requiring models to store and retrieve maximum information content. The benchmark tests single-step and multi-hop induction with order- and position-invariance by shuffling hash pairs.
How much funding has Magic raised?
Magic has raised a total of $515 million. This includes a $5M Seed round, $23M Series A from CapitalG, Nat Friedman, Elad Gil and others, and a recent $320 million investment from Eric Schmidt, Jane Street, Sequoia, Atlassian, plus existing investors Nat Friedman, Daniel Gross, Elad Gil, and CapitalG.
Where are Magic's offices located?
Magic's headquarters is in San Francisco, US. Most job positions are based in SF, with some positions like Member of Technical Staff for Kernels and Supercomputing Platform & Infrastructure also open to remote candidates. The company was founded in 2022 and is privately held with 2-10 employees.
What programming languages does Magic support?
Magic focuses on software development generally rather than specific languages. The system is designed to read entire repositories, documentation, and tickets to plan, write, review, and test code across large codebases. Their prototype model successfully implemented features in open source repos like Documenso and created calculators using custom in-context GUI frameworks.