Microsoft Research Unveils Magnetar: a zkVM 10–39x Faster Than Alternatives

Microsoft Research has introduced Magnetar, a zero-knowledge virtual machine that proves the execution of ordinary programs written in Rust, Go and C. The program runs as usual and, along with its result, produces a short proof that this exact code, on these exact inputs, yielded this exact output. Verification takes about 12 milliseconds. On a laptop, Magnetar processes 7.2–8.7 million CPU instructions per second and on average is 10–39 times faster than the fastest comparable zkVMs. The work's author is Srinath Setty, a senior principal researcher and group lead at Microsoft Research.

Three Problems That Were Never Solved Together

Zero-knowledge proofs were invented in the 1980s, but they never became a general-purpose tool. Three obstacles stood in the way, and each had been overcome only at the expense of the others.

The first is efficiency. Generating a proof used to take millions of times more computation than running the program itself. The second is usability. Many systems force developers to rewrite a program by hand as a cryptographic circuit, which takes months, and any change to the application means a new circuit and new sources of bugs. The universal alternative is to prove the execution of an ordinary processor, but for a long time that cost too much in speed or security.

The third is the cryptographic foundation. Many systems are built on elliptic curves, which a sufficiently powerful quantum computer can break. Post-quantum variants rely on lattices, on new hash functions designed to make proving cheap, or on unproven conjectures about error-correcting codes. Their speed is bought with strong assumptions.

Ten Years of Details: What Magnetar Is Built From

Magnetar is the result of more than a decade of work at Microsoft Research. The researchers proceeded step by step, and each project solved one piece of the puzzle:

• Spartan (2019) revived the sum-check approach and made it practical for general-purpose programs without a trusted setup. It allows different cryptographic building blocks.
• Phalanx (2021) exploits repetition, such as an identical loop of CPU instructions, and proves computations roughly an order of magnitude faster than Spartan. Magnetar's processor description is built on this idea.
• Nova, HyperNova, NeutronNova, Neo and SuperNeo stop a Spartan proof early and fold the partial results together, which makes proving computations of any size significantly cheaper.
• Brakedown (2021) is a fast hash-based commitment scheme. Combined with Spartan, it produced the first variant whose security rests on standard hashes.
• Spice, Lasso, Nebula, Twist and Shout made memory operations cheap, and a general-purpose processor depends on them.
• Jolt showed that, combined with Spartan, the execution of an entire processor can be proved efficiently.
• Vega applied these ideas to digital identity: facts from a government document are proved in less than a tenth of a second.
• BinarySpartan ported Spartan to binary fields, where processor logic and standard hash functions are expressed efficiently.
• Ligerito is a Brakedown-style commitment scheme with noticeably shorter proofs.
• Ring switching lets the system commit to bit vectors efficiently.

Speed and Security in Numbers

Magnetar proves the execution of RISC-V, an open architecture with mature compilers for virtually all major languages. A developer writes a program in Rust, Go or C, compiles it and gets a proof without rewriting anything as circuits. A business policy, a data pipeline or an Ethereum virtual machine interpreter are all ordinary programs.

Key figures:

• 7.2–8.7 million CPU instructions per second on a laptop;
• 9.5–12.8 million cycles per second in the optimized version, which increases all the stated speedups by roughly 1.3x;
• 10–39 times faster on average than the fastest existing zkVMs across 15 test programs, ranging from cryptographic hashing to token transfers in a small EVM interpreter;
• more than 300 times faster than two other systems on the token-transfer test;
• proof verification in about 12 ms.

Magnetar's security rests on a standard hash function such as SHA-256 or SHA-3. It needs no elliptic curves, lattices, specialized hashes, trusted setup or unproven conjectures about codes. Hash functions are the best-studied primitives in cryptography, and hash-based signatures are built on them. NIST has standardized those signatures as a conservative post-quantum option. For Microsoft, this also fits the Quantum Safe program, which is meant to move its products to post-quantum cryptography by 2029.

The architecture itself adds a separate benefit. Verifying a proof consists mainly of hashes and binary-field arithmetic, the very operations Magnetar proves efficiently. Verification therefore becomes an ordinary workload for the system, which can prove computations of any size in pieces and then combine them. That makes it possible to prove long-lived systems such as a ledger or a blockchain.

Where This Is Useful in Practice

The nearest use case for AI is agents. An agent acting on a user's behalf can attach to each action a proof that it complied with the policy it was given. The verifier does not have to trust the agent's operator or read its logs. Regulated financial institutions will be able to prove they processed transactions correctly without revealing the transactions themselves.

For engineers, something else matters more: the barrier to entry is lower. If a team already maintains code in Rust, Go or C, it does not need a cryptography specialist to write circuits. The open question is which results companies will want to verify rather than take on faith. Microsoft Research invites discussion of possible scenarios. The technical paper with the full design, security analysis and measurements is titled "Magnetar: High-speed post-quantum RV64 zkVM from standard hashes".