Magic reveals 100M-token context model and $320M in fresh funding

The biggest bottleneck for AI coding assistants hasn't been intelligence—it has been memory. Feed a model an entire production codebase, and your GPU memory blows up along with your cloud bill.

AI coding startup Magic just unveiled an ultra-long context model named LTM-2-mini, supporting an eye-popping 100 million tokens. Alongside the research, the company announced a fresh $320 million funding round and a massive supercomputing tie-up with Google Cloud and NVIDIA.

100M tokens and the HashHop test

To put 100M tokens in perspective: that is roughly 10 million lines of code or 750 full-length novels. For developers, that means piping your entire repository, private internal libraries, and documentation directly into the inference window without relying on lossy pre-training memorization.

Magic also called out existing long-context benchmarks like Needle In A Haystack as overly simplistic. Dropping a recognizable sentence into irrelevant text lets models cheat using semantic outliers, while state-space models like Mamba rely on special tokens that essentially hand over the answers in advance.

To fix this, Magic open-sourced HashHop. It prompts models with shuffled, uncompressible hash pairs to measure multi-hop reasoning across long sequences—mirroring real-world variable assignments and library imports. Magic's LTM architecture proved it could reliably leap multiple latent hops without breaking down.

Slashing KV cache from 638 H100s to a sliver of one

Context scaling usually dies on the altar of memory bandwidth. Running Meta's Llama 3.1 405B on a 100M token context window would require 638 NVIDIA H100 GPUs per user just to hold the 51TB KV cache.

Magic claims LTM-2-mini requires only a fraction of a single H100's memory for that same 100M context, slashing sequence-dimension decoding compute by roughly 1,000x compared to standard attention.

Even as a downscaled prototype, LTM-2-mini built a functioning calculator inside an unfamiliar custom GUI framework solely from repo context. It also independently shipped a password strength meter for the open-source repo Documenso without human intervention.

Remarkably, Magic pulled this off with just 23 employees. Led by former OpenAI supercomputing lead Ben Chess, the team ditched torch autograd and wrote their entire training and inference stack from scratch in custom CUDA. Backed by 8,000 H100s today, Magic raised $320M from Eric Schmidt, Sequoia, Jane Street, and Atlassian, bringing total capital to $515M.

They are now building the Magic-G5 supercomputer on Google Cloud, powered by NVIDIA GB200 NVL72 racks and designed to scale to tens of thousands of Blackwell GPUs.

Why it matters

We think Magic's move highlights a critical industry shift: pre-training scaling laws are plateauing, and inference-time compute is where real-world agency will be won.

For software engineers, the ultimate value proposition is game-changing. Magic's stated goal is allowing developers to spend $100 and 10 minutes on a bug ticket to reliably receive an end-to-end pull request. While everyday developers can play with HashHop on GitHub right now, the flagship LTM-2 is still cooking on their new cluster.

Magic's smartest play is attacking the memory footprint directly instead of chasing parameter size. If you can handle 100M tokens efficiently, fragmented RAG pipelines suddenly look obsolete for code. However, the catch is real: massive context does not automatically guarantee high-level architectural reasoning, and scaling custom CUDA architectures on tens of thousands of unreleased Blackwell chips is a notoriously brutal engineering hurdle.