Skip to content

openfold3

Protein structure prediction. OpenFold3 (open reimplementation) exposed as HTTP.

Alternatives on the same wire format: Chai-1, Boltz-2. You can swap the image and change nothing else, the engine talks to any of them via the same PREDICT_STRUCTURE_URL env var.

Pre-reqs

  • NVIDIA GPU with ≥40 GB memory for full-length proteins (A100-40G, A100-80G, H100)
  • L40S / A10G work for short sequences (<300 residues)
  • Docker + NVIDIA Container Toolkit
  • CUDA 12.x drivers
  • Model weights (bundled in the image; ~15 GB download on first pull)

Deploy

Via docker-compose:

Uncomment the openfold3 block in docker-compose.yml, then: 

docker compose up openfold3

Standalone:

docker run -d \
  --name novomcp-openfold3 \
  --gpus all \
  -p 8025:8025 \
  -v openfold3-weights:/weights \
  --restart unless-stopped \
  ghcr.io/novomcp/openfold3:latest

First boot pulls the model weights (~15 GB), takes a few minutes.

Wire into the engine

export OPENFOLD3_URL=http://localhost:8025
# or the generic env var, which works for OpenFold3 / Chai / Boltz interchangeably:
export PREDICT_STRUCTURE_URL=http://localhost:8025

Verify

curl -s http://localhost:8025/health
# {"status":"healthy","gpu_available":true,"weights_loaded":true}

Tools that light up

  • predict_structure, from sequence to PDB
  • get_structure_result, retrieve completed predictions

Swapping providers

To use Chai-1 or Boltz instead: 

docker stop novomcp-openfold3
docker run -d --gpus all -p 8025:8025 --name novomcp-chai \
  ghcr.io/novomcp/chai-server:latest
Same env var, same tool call surface.

Env vars

Var Default Purpose
PORT 8025 HTTP listen port
MAX_LENGTH 1500 Reject sequences longer than this (memory guard)
NUM_RECYCLES 3 Recycles per prediction (accuracy vs speed)

Cost

  • Small protein (100 res): ~30 seconds on A100
  • Large protein (1000 res): ~5 minutes on A100-80G
  • Cheapest reliable option: L40S at ~$1/hr spot; A10G struggles above ~400 residues