ML-KEM vs ML-DSA (and SLH-DSA): FIPS 203, 204 and 205 Compared
ML-KEM and ML-DSA are not competing options — they do different jobs, and a post-quantum system needs both. ML-KEM (FIPS 203) is a key-encapsulation mechanism: it lets two parties agree on a shared secret key, replacing ECDH and RSA key exchange. ML-DSA (FIPS 204) is a digital signature algorithm: it proves who signed something, replacing ECDSA and RSA signatures. SLH-DSA (FIPS 205) is the conservative alternative to ML-DSA — not to ML-KEM.
One rule of thumb covers most confusion: KEM = keys, DSA = signatures. NIST published all three standards on August 13, 2024, so the algorithms you migrate to are official, named, and stable.
At a glance
| ML-KEM | ML-DSA | SLH-DSA | |
|---|---|---|---|
| Standard | FIPS 203 | FIPS 204 | FIPS 205 |
| Job | Key establishment | Digital signatures (default) | Digital signatures (conservative) |
| Replaces | ECDH, RSA key transport | ECDSA, RSA signatures | ECDSA, RSA signatures |
| Derived from | CRYSTALS-Kyber | CRYSTALS-Dilithium | SPHINCS+ |
| Security rests on | Module lattices | Module lattices | Hash functions only |
| Middle parameter set | ML-KEM-768 | ML-DSA-65 | SLH-DSA-*-192s / 192f |
| Public key (middle set) | 1,184 bytes | 1,952 bytes | 48 bytes |
| Per-use output (middle set) | 1,088-byte ciphertext | 3,309-byte signature | 16,224 / 35,664-byte signature |
Sizes are from the standards’ own tables: FIPS 203 Table 3, FIPS 204 Table 2, FIPS 205 Table 2.
The two jobs — and what breaks
Public-key cryptography does two things, and a quantum computer breaks today’s tools for both:
- Key establishment — agreeing on a secret key over an open network so traffic can be encrypted. Today: ECDH, RSA. Quantum-broken. Post-quantum replacement: ML-KEM.
- Digital signatures — proving that a message, certificate, or software update is authentic. Today: ECDSA, RSA. Quantum-broken. Post-quantum replacement: ML-DSA, or SLH-DSA where you want hash-only assumptions.
You can’t substitute one for the other. A KEM produces a shared secret; it cannot sign anything. A signature scheme proves authorship; it cannot agree a key. A TLS connection uses both: key exchange for the session keys, signatures for the server’s certificate chain.
ML-KEM (FIPS 203) — key establishment
ML-KEM = Module-Lattice-Based Key-Encapsulation Mechanism, the standardized form of CRYSTALS-Kyber.
- What it’s for: establishing shared secret keys — the quantum-safe replacement for ECDH. This is what protects data in transit against harvest now, decrypt later.
- Parameter sets: ML-KEM-512, ML-KEM-768, and ML-KEM-1024, which FIPS 203 claims meet NIST security categories 1, 3, and 5. All three produce a 32-byte shared secret.
- Where you’ll meet it first: TLS. The hybrid group
X25519MLKEM768(codepoint0x11EC) combines classical X25519 with ML-KEM-768. Chrome began offering it in Chrome 131, and OpenSSL 3.5 offers it in its default key shares. See how hybrid key exchange works in TLS 1.3.
If you migrate one thing first, it’s key establishment — recorded-today traffic is the data most exposed to future decryption.
ML-DSA (FIPS 204) — the default signature
ML-DSA = Module-Lattice-Based Digital Signature Algorithm, the standardized form of CRYSTALS-Dilithium.
- What it’s for: certificates, code signing, authentication tokens — anywhere ECDSA or RSA signs today.
- Parameter sets: ML-DSA-44, ML-DSA-65, ML-DSA-87, claimed at security categories 2, 3, and 5.
- Priority: usually after key establishment. A recorded session can be decrypted retroactively, but a signature can only be forged by an attacker with a quantum computer at the time they use the forgery — so harvest-now risk doesn’t apply to signatures the same way. You still migrate them; NIST’s final transition plan (IR 8547, October 2026) deprecates quantum-vulnerable signature algorithms after 2030 and disallows them after 2035.
SLH-DSA (FIPS 205) — the conservative signature
SLH-DSA = Stateless Hash-Based Digital Signature Algorithm, from SPHINCS+.
- Why it exists: its security rests only on hash functions — a well-studied primitive — rather than on lattice math. If a weakness in lattices ever emerged, ML-KEM and ML-DSA would share it; SLH-DSA would not.
- The cost: tiny keys, large signatures. Each security level comes in an “s” variant (smaller signatures) and an “f” variant (faster signing). The smallest option, SLH-DSA-SHA2-128s, still produces a 7,856-byte signature.
- When to choose it over ML-DSA: long-lived, high-assurance signing where you’d rather not depend on lattice assumptions and size matters less than confidence — firmware, secure boot, root-of-trust keys.
What about FALCON / FN-DSA?
NIST also selected FALCON, to be standardized as FN-DSA (FIPS 206) — a lattice signature scheme with compact signatures. As of October 2026, NIST’s standardization page still lists FIPS 206 as in development, with no publication date. Plan around ML-DSA today.
Worked examples
1. Sizing a TLS handshake (ML-KEM’s cost). With X25519MLKEM768, the client’s key share is 1,216 bytes (1,184 for ML-KEM plus 32 for X25519) and the server’s is 1,120 bytes (1,088 plus 32), per the IETF hybrid ML-KEM draft. A classical X25519 share is 32 bytes. Conclusion: the client share grows about 38×, but it’s paid once per handshake, not per byte of data transferred afterward.
2. Signing a certificate (ML-DSA’s cost). An Ed25519 signature is 64 bytes and its public key 32 bytes (RFC 8032). An ML-DSA-65 signature is 3,309 bytes and its public key 1,952 bytes. A certificate carries the subject’s public key and the issuer’s signature, and a TLS server sends a chain of them. Conclusion: signatures are where post-quantum size really bites, and why signature migration is slower and more carefully staged than key exchange.
3. Failure case — “We enabled ML-KEM, so we’re post-quantum.” A server negotiates X25519MLKEM768, but its certificate is still signed with ECDSA P-256. The session keys are now protected against harvest-now-decrypt-later — a real win. But authentication is still entirely classical. Conclusion: you’ve finished the ML-KEM half, not the migration. Track ML-DSA for certificates and code signing as a separate workstream with its own deadline.
4. Failure case — reaching for SLH-DSA everywhere. A team picks SLH-DSA-SHA2-128f for TLS server certificates “to be safe.” Each signature is 17,088 bytes, more than five times ML-DSA-65’s 3,309, and every handshake carries several. Conclusion: SLH-DSA’s conservatism is worth paying for on rarely-verified, long-lived signatures like firmware; on high-volume handshakes, ML-DSA is the default for a reason.
How to choose — a cheat sheet
| You need to… | Use | Standard |
|---|---|---|
| Establish keys / protect data in transit | ML-KEM | FIPS 203 |
| Sign certificates, code, tokens (default) | ML-DSA | FIPS 204 |
| Sign with the most conservative assumptions (firmware, root of trust) | SLH-DSA | FIPS 205 |
| Sign where signature size is tightly constrained | FN-DSA (not yet final) | FIPS 206 |
What the standards ask of you
The standards define the algorithms; NIST’s transition guidance defines when. The short version: inventory first, migrate key establishment first via hybrids, then signatures. That side is covered in what NIST’s migration guidance actually requires, and the current status of each standard is tracked in NIST post-quantum standards: 2026 status.
From standards to your stack
Knowing the difference between ML-KEM and ML-DSA is useful; knowing where your systems still use their classical predecessors is what an auditor asks for. Start with the surface you can check in a minute: read your own TLS handshake. The negotiated key-exchange group tells you whether ML-KEM is already protecting that connection; the certificate’s signature algorithm tells you how far you are from ML-DSA.
Related: NIST migration guidance · Why post-quantum cryptography exists · NIST post-quantum standards: 2026 status · TLS 1.3 post-quantum
Frequently asked questions
Is ML-KEM symmetric or asymmetric?
Asymmetric. ML-KEM has a public encapsulation key and a private decapsulation key, and its output is a 32-byte shared secret that you then use with a symmetric cipher such as AES to encrypt the actual data.
Is ML-KEM the same as Kyber?
ML-KEM is the standardized algorithm derived from CRYSTALS-Kyber, but the final version includes technical changes that make it incompatible with the earlier deployed Kyber. In TLS that meant a new codepoint: hybrid Kyber768 used 0x6399, while hybrid ML-KEM uses 0x11EC.
What is ML-KEM-768?
It is the middle of ML-KEM's three parameter sets, claimed by NIST to meet security category 3. It uses a 1,184-byte encapsulation key and a 1,088-byte ciphertext, and it is the ML-KEM half of the X25519MLKEM768 hybrid used in TLS.
What is ML-DSA-65?
It is the middle of ML-DSA's three parameter sets, claimed to meet security category 3, with a 1,952-byte public key and a 3,309-byte signature. ML-DSA-44 is smaller (category 2) and ML-DSA-87 is larger (category 5).
Does OpenSSL support ML-KEM and ML-DSA?
Yes. OpenSSL 3.5 added ML-KEM, ML-DSA and SLH-DSA, and changed its default TLS key shares to offer X25519MLKEM768 alongside X25519.