Proof-based general-purpose computing with Magnetar

Microsoft Research says it has pushed zero-knowledge proofs a lot closer to everyday computing with Magnetar, a new zero-knowledge virtual machine for ordinary code. The idea is simple enough to say out loud: a program runs, and it hands back not just an output but a short cryptographic proof that the output is correct. Anyone can check that proof quickly, without re-executing the workload and without trusting the machine that produced it.

What makes Magnetar stand out is that it tackles three long-running problems at once. It is fast enough to matter, general enough for normal development workflows, and built on standard hash functions rather than the newer cryptographic tricks that have powered many faster systems. Microsoft says it proves Rust, Go, and C programs as written, because it targets RISC-V and works with mainstream compilers. In other words, engineers do not have to translate applications into a custom circuit first, which has been one of the more painful bottlenecks in this field.

The performance claims are strong. On a laptop, Magnetar proves 7.2 to 8.7 million processor instructions per second, and an optimized version reaches 9.5 to 12.8 million cycles per second. Across 15 benchmark programs, it is 10 to 39 times faster on average than the fastest existing zkVMs on the same machine, and more than 300 times faster than two other systems on a token-transfer benchmark. Its proofs verify in about 12 milliseconds.

The other big shift is cryptographic caution. Magnetar does not rely on elliptic curves, lattices, special-purpose hash functions, or a trusted setup. Microsoft says that matters because hash functions like SHA-256 and SHA-3 are already widely trusted, and the company wants systems that can still be trusted decades from now. That is a pretty sober answer to a field that has often chased speed first and asked about the maintenance bill later.

Magnetar also builds on more than a decade of Microsoft Research work, including Spartan, Phalanx, Brakedown, Jolt, BinarySpartan, and others. This is less a sudden leap than a pile-up of pieces finally fitting together. But the result is still notable: proof-based computing is starting to look less like a research demo and more like something a real system could use.