Models & limits¶

The live, machine-readable version of this page is GET /v1/models. It returns the model list, every parameter with its default/range, the design protocols, and the current numeric limits. The tables below mirror it; if they ever disagree, trust the live endpoint.

curl -s https://api.japanfold.com/v1/models

GET /v1/health is a plain liveness check ({"status":"ok","api_version":"…"}) if you just need to know the service is up.

Models¶

`id`	Name	MSA	Ligands	Nucleic acids	Affinity	Constraints	PAE/PDE
`boltz2`	Boltz-2	default on	✓	✓	✓	✓	-
`esmfold2`	ESMFold-2	optional	-	-	-	-	-
`esmfold2-fast`	ESMFold-2 Fast	single-seq only	-	-	-	-	-
`protenix-v2`	Protenix-v2	default on	✓	✓	-	-	✓

Boltz-2: most capable; choose it when in doubt, or whenever you need ligands, nucleic acids, affinity or constraints.
ESMFold-2 / ESMFold-2 Fast: language-model folding, protein chains only. Fast, lightweight; -fast is always single-sequence, for screening many sequences at once.
Protenix-v2: AlphaFold3-family (Pairformer + atom diffusion), strong at antibody–antigen. Emits PAE/PDE and per-atom pLDDT. For affinity, use Boltz-2.

Prediction parameters¶

Sent as params on POST /v1/predictions. Out-of-range values are clamped.

Key	Type	Default	Range	Notes
`use_msa_server`	bool	`true`	-	Build an MSA. Required for Boltz-2/Protenix; optional for ESMFold-2.
`fast`	bool	`true`	-	Higher throughput, may be slightly less accurate.
`recycling_steps`	int	`3`	1–10	More can improve accuracy, slower.
`sampling_steps`	int	`200`	10–500	Diffusion steps per structure.
`diffusion_samples`	int	`1`	1–5	Structures generated per target.
`output_format`	enum	`cif`	`cif`, `pdb`	Structure file format.

Design protocols & parameters¶

Protocols for POST /v1/designs: protein-anything, peptide-anything, nanobody-anything, antibody-anything, protein-small_molecule, protein-redesign. See Designs for what each does.

Key	Type	Default	Range	Notes
`num_designs`	int	`10`	1–10	Binders to generate before filtering.
`budget`	int	`10`	1–10	Top ranked designs to keep after filtering.
`fast`	bool	`true`	-	Higher throughput, may be slightly less accurate.

Limits (free tier)¶

This is a free public demo on shared compute, so inputs and concurrency are capped. The full platform has no such limits.

Per structure / complex

Limit	Value
`max_residues` (per structure)	1024
`max_chains_per_complex`	10
`max_ligands_per_complex`	10
`max_constraints_per_complex`	20
`max_complexes` (structures per run)	10
`max_content_chars` (per input string)	50000

Design

Limit	Value
`max_designs`	10
`max_budget`	10

Parameter ceilings

Limit	Value
`max_recycling_steps`	10
`max_sampling_steps`	500
`max_diffusion_samples`	5

Concurrency & rate

Limit	Value
`max_active_jobs` (service-wide)	64
`max_active_jobs_per_ip`	8
`max_active_jobs_per_session`	3
`max_submits_per_min` (service-wide)	12
`max_submits_per_min_per_ip`	40
`max_retained_jobs`	200

Runtime / stall guards

Limit	Value
`max_runtime_predict_s`	1500
`max_runtime_design_s`	2700
`max_stall_s` (predict)	600
`max_stall_design_s`	1200

Exceed a size cap → 400. At capacity / over a rate limit → 429 with Retry-After. See Errors.