Cog
Containers for ML: package models without the Dockerfile pain.
About
Cog is Replicate's open-source tool for packaging ML models as production-ready Docker containers: define the environment in a simple cog.yaml (it resolves the CUDA/cuDNN/PyTorch combos for you — 'no more CUDA hell'), define inputs/outputs in Python, and Cog generates the image plus an automatic HTTP inference server with an OpenAPI schema. 9.5k stars, Apache-2.0. Deploy the image anywhere Docker runs, or push to Replicate. It's packaging, not hosting — you still need the GPUs — but it kills the most annoying part of model deployment.
What it can do
- **No CUDA hell**: compatible CUDA/cuDNN/PyTorch combos resolved for you
- **Auto API server**: Rust/Axum HTTP server from your Python types
- **OpenAPI schema**: typed inputs/outputs with validation
- **Reproducible**: lockfile-style builds with caching
- **Deploy anywhere**: your infra or Replicate
Setup
```bash
# install per replicate.github.io/cog (binary or Docker)
cog predict -i prompt="a photo of a cat" # local test
cog push r8.im/yourname/yourmodel # push to Replicate
```
Define cog.yaml + predict.py. Docs: replicate.github.io/cog.
Prerequisites
Docker; NVIDIA GPU for GPU models. Replicate account for `cog push`.
Use cases
Works with
Trust score — 9/10
Scored 9/10 (verified tier, metrics-v1). Evidence: 9,489 GitHub stars (log-scaled vs catalog peers); blended with ~245,827 monthly package downloads; last push 2d ago; 28 releases/12mo; 88 open issues. Manual review: Replicate's model packaging (no CUDA hell, auto OpenAPI server), 9.5k stars; packaging not hosting — you still need GPUs. License: Apache-2.0.
★ 9,489 GitHub stars
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.7/10
- Maintenance · recency, releases, issue load
- 9.6/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · 9,489 GitHub stars (log-scaled vs catalog peers)
- · blended with ~245,827 monthly package downloads
- · last push 2d ago; 28 releases/12mo; 88 open issues
- · manual review: Replicate's model packaging (no CUDA hell, auto OpenAPI server), 9.5k stars; packaging not hosting — you still need GPUs
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.