Same prefix, different model.
Three pairs that get confused in conversation, laid out row by row. The reading under each table says what the difference amounts to once the specification sheet is out of the way.
ESM-2, ESM3 and ESM C
Three models with the same prefix that answer different questions.| ESM-2 | ESM3 | ESM C | |
|---|---|---|---|
| What it is | Masked sequence encoder | Multimodal generative model | Masked sequence encoder |
| Tracks | Sequence | Sequence, structure, function | Sequence |
| Generates | Masked-token logits | Sequence, structure and annotation | Masked-token logits |
| Open sizes | 8M to 15B | 1.4B open checkpoint | 300M, 600M, 6B |
| Context | 1024 tokens | See model card | 2048 tokens |
| Use it for | Embeddings and baselines | Prompted design | Embeddings at larger scale |
| Published as | Science 2023 | Science 2025 | Preprint 2026 |
ESM3 is the generative line and ESM C the representation line. ESM-2 remains the reference point both are measured against, and it is still the checkpoint most downstream code loads.
AlphaFold 2 and AlphaFold 3
What actually changed between the two.| AlphaFold 2 | AlphaFold 3 | |
|---|---|---|
| Trunk | Evoformer, MSA and pair streams | Pairformer, token and pair only |
| Alignment | Carried through the trunk | Summarised early, then set aside |
| Decoder | Structure module, residue frames | Diffusion over atoms |
| Output | One structure with confidence | Sampled structures, then ranked |
| Covers | Proteins, and multimers by extension | Proteins, nucleic acids, ligands, ions |
| Triangle updates | Yes | Yes |
| Published as | Nature 2021 | Nature 2024 |
The triangle machinery survived and the alignment stream did not. Read the difference as a move from one deterministic answer about proteins to a sampled answer about biomolecular complexes.
ESMFold and ESMFold2
The same idea, rebuilt on a newer encoder and a different decoder.| ESMFold | ESMFold2 | |
|---|---|---|
| Encoder | ESM-2, frozen | ESM C |
| Alignment | None at query time | Optional, for difficult targets |
| Decoder | AlphaFold2-style structure module | All-atom diffusion |
| Scope | Single chains | Chains and complexes |
| Design use | Prediction only | Inversion used for binder design |
| Status | Science 2023 | Preprint 2026, not peer reviewed |
The lineage is a language model feeding a folding decoder. What changed is that the decoder became generative and the alignment came back as an option rather than a requirement.